refactor(pf-encode): split vulkan_video's construction tail into vk_build.rs (WP7.5)

The ~820-line tail of free builders — make_frame/make_frame_csc/ make_frame_common, make_video_image, probe_rgb_direct, the H.265/AV1 parameter-set writers and the AV1 bit-writer — moves to a #[path] child module, the amf_sys.rs shape: the child sees the parent's private items (Frame and friends), so the split costs zero visibility churn. Six parent-called items went pub(super); five stay child-private (dead_code is per-item and each is used within the child). vulkan_video.rs drops 5,292 → 4,489 lines and the construction unsafe gets its own review surface; steady-state encode logic stays in the parent. ⚠ Trap recorded for future child-module splits: inline `use super::X` statements INSIDE moved fn bodies silently change meaning (super shifts one level) — vk_av1_encode/vk_valve_rgb imports needed crate:: paths. Proven on-glass, not just compiled: all 8 vulkan GPU smokes green under the validation layers on the 780M post-split (H.265 + AV1, RGB-direct, CSC, CPU paths — every moved constructor exercised). nvenc_cuda.rs and qsv.rs are DECLINED the same treatment, with evidence in the handoff doc: no equivalent self-contained seam — their candidate regions are ~150-line loader/accessor clusters interleaved with the encoders' own state types, and a thin-forwarder impl split is exactly the churn a no-defect phase penalizes. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
refactor(pf-encode): one Linux backend resolver, consumed by dispatch AND mirrors (WP7.6)
2026-07-26 00:08:03 +02:00 · 2026-07-26 00:08:03 +02:00 · 2026-07-26 00:08:03 +02:00 · 2026-07-26 00:08:03 +02:00 · 2026-07-26 00:08:03 +02:00 · 2026-07-26 00:08:03 +02:00
88 changed files with 9259 additions and 3145 deletions
@@ -1,9 +1,12 @@
-# Root build context is used only by web/Dockerfile, which needs web/ and
-# api/openapi.json. Allowlist those; keep everything else (target/, .git, crates)
-# out of the context upload.
+# The root build context is used by web/Dockerfile (which needs web/ and
+# api/openapi.json) and by ci/rust-ci-arm64cross.Dockerfile (which needs the toolchain
+# pin). Allowlist those; keep everything else (target/, .git, crates) out of the
+# context upload.
 *
 !web
 !api/openapi.json
+!rust-toolchain.toml
+!ci/pf-host-cc
 web/node_modules
 web/.output
 web/dist
@@ -61,6 +61,43 @@ jobs:
      - name: Test (unit + loopback + proptest + C ABI harness)
        run: cargo test --workspace --locked

+      # The GPU encode backends are OFF by default, so every step above compiles ~none of them:
+      # `nvenc` gates enc/linux/nvenc_cuda.rs (+ nvenc_core/nvenc_status) and `vulkan-encode` gates
+      # enc/linux/vulkan_video.rs (+ the vendored vk_av1_encode/vk_valve_rgb bindings) — ~8,150
+      # lines carrying ~70 `unsafe` blocks. Their ONLY prior CI coverage was deb.yml's
+      # `cargo build`, where warnings are not errors, so pf-encode's own
+      # `#![deny(clippy::undocumented_unsafe_blocks)]` — the crate's stated unsafe-proof gate —
+      # was never actually enforced on them. (`pyrowave` needs no extra step: punktfunk-host has
+      # `default = ["pyrowave"]`, so the steps above already cover it.)
+      #
+      # `--all-targets` is load-bearing, not decoration: without it the feature-gated
+      # `#[cfg(test)]` modules are never compiled, which is exactly how all ten
+      # `NvencCudaEncoder::open` call sites in nvenc_cuda.rs's tests drifted to the wrong arity
+      # (E0061 x10) without any job noticing.
+      #
+      # GPU-free: every test needing real hardware is `#[ignore]`d, and NVENC/CUDA resolve their
+      # entry points at RUNTIME (dlopen), so the test binary links without a driver present.
+      # (On MSVC the same crate link-imports those symbols instead, which is why windows-host.yml
+      # can only type-check these tests via clippy — see the note there.)
+      #
+      # Scoped to `-p pf-encode` with ITS OWN feature names: punktfunk-host has no code gated on
+      # `nvenc`/`vulkan-encode` (its only `cfg(feature)` sites are the two `pyrowave` ones in
+      # capture.rs, and pyrowave is default-on, so the steps above already cover them). Going
+      # through `--features punktfunk-host/...` would force punktfunk-host into the selection and
+      # re-run its entire test suite a second time for no extra coverage.
+      #
+      # `pyrowave` is listed explicitly even though it is punktfunk-host's default: selecting only
+      # `-p pf-encode` takes the host out of the resolution, and pf-encode's own default is empty.
+      # Naming it keeps this the SHIPPED Linux feature set — deb.yml builds
+      # `--features punktfunk-host/nvenc,punktfunk-host/vulkan-encode` WITHOUT
+      # `--no-default-features`, so the .deb carries nvenc + vulkan-encode + pyrowave together, and
+      # that combination is what deserves the lint.
+      - name: Clippy + test the feature-gated Linux encode backends
+        run: |
+          cargo clippy -p pf-encode --all-targets --locked \
+            --features nvenc,vulkan-encode,pyrowave -- -D warnings
+          cargo test -p pf-encode --locked --features nvenc,vulkan-encode,pyrowave
+
      - name: C ABI harness (standalone link proof)
        run: bash crates/punktfunk-core/tests/c/run.sh

@@ -71,6 +108,53 @@ jobs:
          git diff --exit-code include/punktfunk_core.h \
            || (echo "include/punktfunk_core.h is stale — commit the regenerated header" && exit 1)

+  # The client stack cross-checked for aarch64. NOT an artifact job — deb.yml ships those —
+  # this exists so a portability defect fails here instead of surfacing in a release build or
+  # on a user's board. It earns its runtime: the bug that motivated it (a Vulkan extension
+  # array typed `*const i8`, where `c_char` is signed on x86_64 and UNSIGNED on aarch64)
+  # compiled cleanly on every target CI built at the time.
+  #
+  # Client crates only, listed explicitly: the host's encode stack is x86 (NVENC/QSV/AMF) and
+  # `--workspace` would drag it in. Runs in the cross image (amd64 toolchain + arm64 sysroot,
+  # ci/rust-ci-arm64cross.Dockerfile) on the ordinary runner — no arm64 runner involved.
+  rust-arm64:
+    runs-on: ubuntu-24.04
+    container:
+      image: git.unom.io/unom/punktfunk-rust-ci-arm64cross:latest
+    timeout-minutes: 60
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Cache keys
+        run: echo "rustc=$(rustc --version | cut -d' ' -f2)" >> "$GITHUB_ENV"
+      - uses: actions/cache@v4
+        with:
+          path: |
+            /usr/local/cargo/registry
+            /usr/local/cargo/git
+          key: cargo-home-${{ hashFiles('Cargo.lock') }}
+          restore-keys: cargo-home-
+      - uses: actions/cache@v4
+        with:
+          path: target
+          # Its OWN prefix: aarch64 artifacts must never share the amd64 jobs' target cache.
+          key: cargo-target-arm64-v1-${{ env.rustc }}-${{ hashFiles('Cargo.lock') }}
+          restore-keys: cargo-target-arm64-v1-${{ env.rustc }}-
+
+      - name: Clippy for aarch64 (deny warnings)
+        run: |
+          cargo clippy --target aarch64-unknown-linux-gnu --all-targets --locked \
+            -p punktfunk-core -p pf-client-core -p pf-presenter -p pf-console-ui \
+            -p punktfunk-client-session -p punktfunk-client-linux \
+            -- -D warnings
+
+      # The minimal embedded build — no Skia, no PyroWave — is what a small image installs, so
+      # it has to keep compiling on its own, not just as a subset of the default features.
+      - name: Build the session binary, minimal features
+        run: |
+          cargo build --release --target aarch64-unknown-linux-gnu --locked \
+            -p punktfunk-client-session --no-default-features
+
  web:
    runs-on: ubuntu-24.04
    container:
@@ -1,9 +1,13 @@
 # Build the punktfunk .debs and publish them to Gitea's Debian package registry, so Ubuntu
-# boxes get new builds via `apt update && apt upgrade`. Two jobs, both publishing to the same
+# boxes get new builds via `apt update && apt upgrade`. Three jobs, all publishing to the same
 # apt distribution/component:
 #
 #   build-publish       — client + web + scripting, on the Ubuntu 26.04 rust-ci image (the client
 #                         needs 24.04-absent libs: SDL3, GTK4 ≥ 4.20).
+#   build-publish-client-arm64
+#                       — the same client package for arm64, CROSS-compiled on the same amd64
+#                         runner in the rust-ci-arm64cross image. No host counterpart: the Linux
+#                         host's encode stack is x86 (NVENC/QSV/AMF).
 #   build-publish-host  — the HOST, on the Ubuntu 24.04 rust-ci-noble image with a from-source
 #                         FFmpeg 8 BUNDLED into the .deb. This lowers the host's glibc floor to 2.39
 #                         and removes the hard `Depends: libavcodec62`, so the ONE host .deb installs
@@ -269,3 +273,94 @@ jobs:
          for DEB in dist/*.deb; do
            upsert_asset "$RID" "$DEB"
          done
+
+  # ---------------------------------------------------------------------------------------------
+  # The aarch64 CLIENT .deb. Cross-compiled on the ordinary amd64 runner in the
+  # punktfunk-rust-ci-arm64cross image (the rust-ci toolchain + an arm64 multiarch sysroot — see
+  # ci/rust-ci-arm64cross.Dockerfile); there is no arm64 runner in the fleet and none is needed.
+  # Client only, by decision: the Linux host encodes with NVENC/QSV/AMF, all x86.
+  # Publishes to the same distribution/component as the amd64 jobs — the apt registry keys pool
+  # entries by arch, so `apt` on an arm64 box picks this one up with no client-side configuration.
+  build-publish-client-arm64:
+    runs-on: ubuntu-24.04
+    container:
+      image: git.unom.io/unom/punktfunk-rust-ci-arm64cross:latest
+    timeout-minutes: 90
+    steps:
+      - uses: actions/checkout@v4
+
+      # Byte-identical to build-publish's version step (pf-version.sh is deterministic per
+      # commit), so the arm64 package always shares the amd64 version line.
+      - name: Version + channel
+        run: |
+          eval "$(bash scripts/ci/pf-version.sh)"
+          SHORT=$(echo "$GITHUB_SHA" | cut -c1-8)
+          case "$GITHUB_REF" in
+            refs/tags/v*) V="${GITHUB_REF_NAME#v}"; DIST=stable ;;
+            *)            V="${PF_BASE}~ci${GITHUB_RUN_NUMBER}.g${SHORT}"; DIST=canary ;;
+          esac
+          echo "VERSION=$V" >> "$GITHUB_ENV"
+          echo "DISTRIBUTION=$DIST" >> "$GITHUB_ENV"
+          echo "package version $V -> apt distribution '$DIST' (arm64)"
+
+      # dpkg-shlibdeps + dpkg-deb. The arm64 link deps themselves are the cross image's whole
+      # point and are already baked in; python3 is for scripts/ci/gitea-release.sh.
+      - name: dpkg-dev
+        run: |
+          apt-get update
+          apt-get install -y --no-install-recommends dpkg-dev python3
+
+      - name: Cache keys
+        run: echo "rustc=$(rustc --version | cut -d' ' -f2)" >> "$GITHUB_ENV"
+      - uses: actions/cache@v4
+        with:
+          path: |
+            /usr/local/cargo/registry
+            /usr/local/cargo/git
+          key: cargo-home-${{ hashFiles('Cargo.lock') }}
+          restore-keys: cargo-home-
+      - uses: actions/cache@v4
+        with:
+          path: target
+          # Its OWN key — these are aarch64 artifacts under target/aarch64-unknown-linux-gnu/
+          # and must never share the amd64 jobs' target cache.
+          key: cargo-target-arm64-v1-${{ env.rustc }}-${{ hashFiles('Cargo.lock') }}
+          restore-keys: cargo-target-arm64-v1-${{ env.rustc }}-
+
+      - name: Build the arm64 client .deb
+        env:
+          PUNKTFUNK_BUILD_VERSION: ${{ env.VERSION }}   # stamped into the binaries (build.rs)
+        run: |
+          git config --global --add safe.directory "$PWD"
+          ARCH=arm64 TARGET=aarch64-unknown-linux-gnu \
+            bash packaging/debian/build-client-deb.sh
+          # Fail here rather than shipping an amd64 binary under an arm64 package name.
+          readelf -h target/aarch64-unknown-linux-gnu/release/punktfunk-session \
+            | grep -q AArch64 || { echo "ERROR: session binary is not AArch64"; exit 1; }
+
+      - name: Publish to the Gitea apt registry
+        env:
+          TOKEN: ${{ secrets.REGISTRY_TOKEN }}
+        run: |
+          for DEB in dist/*.deb; do
+            echo "uploading $DEB"
+            NAME=$(dpkg-deb -f "$DEB" Package)
+            VER=$(dpkg-deb -f "$DEB" Version)
+            ARCH=$(dpkg-deb -f "$DEB" Architecture)
+            curl -fsS -o /dev/null --user "enricobuehler:$TOKEN" -X DELETE \
+              "https://$REGISTRY/api/packages/$OWNER/debian/pool/$DISTRIBUTION/$COMPONENT/$NAME/$VER/$ARCH" || true
+            curl -fsS --user "enricobuehler:$TOKEN" --upload-file "$DEB" \
+              "https://$REGISTRY/api/packages/$OWNER/debian/pool/$DISTRIBUTION/$COMPONENT/upload"
+          done
+          echo "published arm64 client to $OWNER/debian $DISTRIBUTION/$COMPONENT"
+
+      - name: Attach the arm64 .deb to the Gitea release (stable tags only)
+        if: startsWith(gitea.ref, 'refs/tags/v')
+        env:
+          GITEA_TOKEN: ${{ secrets.REGISTRY_TOKEN }}
+        run: |
+          . scripts/ci/gitea-release.sh
+          RID=$(ensure_release "$GITHUB_REF_NAME" "$GITHUB_REF_NAME" auto)
+          for DEB in dist/*.deb; do
+            upsert_asset "$RID" "$DEB"
+          done
@@ -2,6 +2,7 @@
 #   punktfunk-web      — management console (web/Dockerfile, repo-root context)
 #   punktfunk-docs     — documentation site (docs-site/Dockerfile)
 #   punktfunk-rust-ci  — Rust CI builder image consumed by ci.yml
+#   punktfunk-rust-ci-arm64cross — the above + an arm64 sysroot, for the aarch64 client legs
 #   punktfunk-fedora-rpm — Fedora 43 builder image consumed by rpm.yml (Bazzite RPM)
 # Host and clients are intentionally NOT containerized (see CLAUDE.md "What's left").
 #
@@ -80,6 +81,41 @@ jobs:
          docker push "$REGISTRY/$OWNER/${{ matrix.image }}:latest"
          case "$GITHUB_REF" in refs/tags/v*) docker push "$REGISTRY/$OWNER/${{ matrix.image }}:${GITHUB_REF_NAME}" ;; esac

+  # The aarch64 CROSS builder — a SEPARATE job because it is `FROM punktfunk-rust-ci:latest`
+  # and so must not race the matrix entry that publishes that base. Consumed by the arm64
+  # client legs in deb.yml/rpm.yml/arch.yml. Root context: it needs rust-toolchain.toml to
+  # install the target against the toolchain the workspace actually pins.
+  build-push-arm64cross:
+    runs-on: ubuntu-24.04
+    needs: build-push
+    timeout-minutes: 45
+    env:
+      IMAGE: punktfunk-rust-ci-arm64cross
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Login to registry
+        run: |
+          echo "${{ secrets.REGISTRY_TOKEN }}" \
+            | docker login "$REGISTRY" -u enricobuehler --password-stdin
+
+      - name: Build
+        run: |
+          EXTRA=""
+          case "$GITHUB_REF" in refs/tags/v*) EXTRA="-t $REGISTRY/$OWNER/$IMAGE:${GITHUB_REF_NAME}" ;; esac
+          docker build --pull \
+            -f ci/rust-ci-arm64cross.Dockerfile \
+            -t "$REGISTRY/$OWNER/$IMAGE:latest" \
+            -t "$REGISTRY/$OWNER/$IMAGE:sha-${GITHUB_SHA::8}" \
+            $EXTRA \
+            .
+
+      - name: Push
+        run: |
+          docker push "$REGISTRY/$OWNER/$IMAGE:sha-${GITHUB_SHA::8}"
+          docker push "$REGISTRY/$OWNER/$IMAGE:latest"
+          case "$GITHUB_REF" in refs/tags/v*) docker push "$REGISTRY/$OWNER/$IMAGE:${GITHUB_REF_NAME}" ;; esac
+
  # Deploy the docs site to unom-1, the DMZ services VM website/cms also deploy to
  # (docs.punktfunk.unom.io via Caddy on home-reverse-proxy-1 -> :3220). Same secret set
  # as unom/website's deploy: DEPLOY_HOST/DEPLOY_USER/DEPLOY_PORT/DEPLOY_SSH_KEY (the
@@ -50,6 +50,23 @@ on:
      # builds — without these, encoder changes only reached this workflow via Cargo.lock luck.
      - 'crates/pf-encode/**'
      - 'crates/libvpl-sys/**'
+      # …and the rest of the W6 subsystem crates this build compiles. pf-encode was listed while
+      # the crates it speaks (pf-frame's CapturedFrame/PixelFormat/dxgi vocabulary, pf-gpu's
+      # adapter selection, pf-zerocopy, pf-host-config) were not, so a change that broke the
+      # Windows host through one of THEM reached main with no Windows build at all — the same
+      # Cargo.lock-luck gap the two lines above were added to close.
+      - 'crates/pf-frame/**'
+      - 'crates/pf-gpu/**'
+      - 'crates/pf-zerocopy/**'
+      - 'crates/pf-host-config/**'
+      - 'crates/pf-capture/**'
+      - 'crates/pf-win-display/**'
+      - 'crates/pf-vdisplay/**'
+      - 'crates/pf-inject/**'
+      - 'crates/pf-paths/**'
+      - 'crates/pf-driver-proto/**'
+      - 'crates/pf-clipboard/**'
+      - 'crates/pyrowave-sys/**'
      - 'packaging/windows/**'
      - 'scripts/windows/**'
      - 'web/**'
@@ -154,8 +171,23 @@ jobs:
        # build minutes earlier). Linting in release reuses those native build-script artifacts (no
        # openh264 rebuild), and keeps everything in one C:\t\release tree. Same reason
        # pf-vkhdr-layer's clippy below runs --release.
+        #
+        # pf-encode is linted SEPARATELY with --all-targets so its Windows `#[cfg(test)]` modules
+        # are type-checked — the AMF C-ABI layout assertions (`variant_layout_matches_c` and
+        # friends, which are the only guard on a hand-mirrored vtable ABI), the QSV tests, and the
+        # PyroWave-Windows smoke test. The host lint above cannot cover them: `-p punktfunk-host`
+        # only builds pf-encode as a dependency, so its test targets are never compiled, and that
+        # blind spot is what let the Linux twin's tests rot to the wrong arity unnoticed.
+        # NOTE: clippy (a check, no link step) is deliberately the vehicle here — `cargo test`
+        # with `nvenc` cannot LINK on MSVC: nvidia-video-codec-sdk link-imports
+        # NvEncodeAPICreateInstance / NvEncodeAPIGetMaxSupportedVersion, which resolve only against
+        # the driver's import lib. (On Linux the same crate dlopens them, so ci.yml can and does
+        # run the tests there.) Running them here would need an `--features amf-qsv,qsv` build
+        # without `nvenc`, i.e. a third full dep tree on a runner that already trips C1069 — not
+        # worth it while ci.yml executes the same tests.
        run: |
          cargo clippy --release -p punktfunk-host --features nvenc,amf-qsv,qsv -- -D warnings; if ($LASTEXITCODE) { throw "host clippy" }
+          cargo clippy --release -p pf-encode --all-targets --features nvenc,amf-qsv,qsv -- -D warnings; if ($LASTEXITCODE) { throw "pf-encode clippy" }
          cargo clippy --release -p punktfunk-tray -- -D warnings; if ($LASTEXITCODE) { throw "tray clippy" }

      - name: Build + lint the HDR Vulkan layer (pf-vkhdr-layer)
@@ -3073,6 +3073,7 @@ dependencies = [
 "libloading",
 "serde",
 "serde_json",
+ "tempfile",
 "tracing",
 ]

@@ -5106,7 +5106,7 @@
        "properties": {
          "audio_streaming": {
            "type": "boolean",
-            "description": "True while the audio stream thread is running."
+            "description": "True while audio is streaming on either plane (same rule as `video_streaming`)."
          },
          "conflicts": {
            "type": "array",
@@ -5150,7 +5150,7 @@
              },
              {
                "$ref": "#/components/schemas/SessionInfo",
-                "description": "The active launch session (set by Moonlight's `/launch`, cleared on cancel/stop)."
+                "description": "The active session: GameStream's launch (Moonlight `/launch`) when present, else the first\nlive native session. `null` when nothing is streaming."
              }
            ]
          },
@@ -5160,7 +5160,7 @@
          },
          "video_streaming": {
            "type": "boolean",
-            "description": "True while the video stream thread is running."
+            "description": "True while video is streaming on EITHER plane: the GameStream media pipeline, or a live\nnative (punktfunk/1) session — the default plane, invisible in the GameStream flag alone."
          }
        }
      },
@@ -0,0 +1,45 @@
+#!/bin/bash
+# Host-side C compiler wrapper for the aarch64 cross image (ci/rust-ci-arm64cross.Dockerfile).
+#
+# Why this exists: ffmpeg-sys-next's build script compiles a probe it intends to RUN — it
+# executes the binary to read the libav* version macros — so it forces `.target(HOST)` with
+# the comment "don't cross-compile this", but still hands that host compile the TARGET's
+# pkg-config include paths. `-I/usr/include/aarch64-linux-gnu` then shadows the host's own
+# multiarch libc headers and the x86 compiler dies inside bits/math-vector.h on NEON/SVE
+# types it has never heard of.
+#
+# Prepending the host's multiarch dir does NOT fix it: GCC drops a `-I` that duplicates a
+# directory already on its system include path (keeping it in the original, later position),
+# so the arm64 dir stays in front. The reliable fix is to remove the target include dirs from
+# the host compile entirely — the probe only wants FFmpeg's version macros, and the amd64
+# libav*-dev headers are installed and on the default search path, at the same version (both
+# come from this Ubuntu release).
+#
+# Scope: only ever invoked as CC for the HOST triple (CC_x86_64_unknown_linux_gnu). Target
+# compiles go to aarch64-linux-gnu-gcc and never pass through here.
+set -euo pipefail
+
+declare -a out=()
+while (($#)); do
+  case "$1" in
+  # `-I dir` as two arguments — the form cc's Command building and ffmpeg-sys both emit.
+  -I)
+    if [[ ${2-} == *aarch64-linux-gnu* ]]; then
+      shift 2
+      continue
+    fi
+    out+=("$1" "${2-}")
+    shift 2
+    ;;
+  # `-Idir` glued into one argument.
+  -I*aarch64-linux-gnu*)
+    shift
+    ;;
+  *)
+    out+=("$1")
+    shift
+    ;;
+  esac
+done
+
+exec /usr/bin/cc "${out[@]}"
@@ -0,0 +1,81 @@
+# Cross-compiling CI builder: amd64 host toolchain + an arm64 multiarch sysroot, for the
+# aarch64 Linux CLIENT artifacts (punktfunk-client + punktfunk-session).
+#
+#   docker build -f ci/rust-ci-arm64cross.Dockerfile -t punktfunk-rust-ci-arm64cross .
+#
+# Derived from punktfunk-rust-ci so the Rust toolchain, clang, and CMake are byte-identical
+# to the amd64 legs — this image only adds the target side. Kept as a SEPARATE image rather
+# than folded into the base because the :arm64 dev libs are ~1 GB that every other CI job
+# would otherwise pull for nothing.
+#
+# Client only: the Linux HOST stays amd64 (its encode stack is NVENC/QSV/AMF), so none of the
+# host's CUDA/GBM link deps are mirrored here.
+#
+# Ubuntu splits archives by architecture: amd64 lives on archive.ubuntu.com, every port
+# (arm64 included) on ports.ubuntu.com. Both stanzas therefore have to be pinned with an
+# explicit `Architectures:` or apt tries to fetch arm64 from the amd64 mirror and 404s.
+#
+# Built from the REPO ROOT context (not ci/) — see the rust-toolchain.toml copy below.
+FROM git.unom.io/unom/punktfunk-rust-ci:latest
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+# 1. Pin the stock sources to amd64, add ports.ubuntu.com for arm64.
+RUN sed -i 's|^Types: deb$|Types: deb\nArchitectures: amd64|' /etc/apt/sources.list.d/ubuntu.sources \
+    && . /etc/os-release \
+    && printf 'Types: deb\nArchitectures: arm64\nURIs: http://ports.ubuntu.com/ubuntu-ports/\nSuites: %s %s-updates %s-backports %s-security\nComponents: main universe restricted multiverse\nSigned-By: /usr/share/keyrings/ubuntu-archive-keyring.gpg\n' \
+        "$VERSION_CODENAME" "$VERSION_CODENAME" "$VERSION_CODENAME" "$VERSION_CODENAME" \
+        > /etc/apt/sources.list.d/ubuntu-ports-arm64.sources \
+    && dpkg --add-architecture arm64
+
+# 2. The cross toolchain + every arm64 dev lib the client links. Mirrors the client half of
+#    rust-ci.Dockerfile's list (FFmpeg, PipeWire, Opus, SDL3, GTK4/libadwaita, xkbcommon,
+#    Vulkan headers for pf-ffvk's bindgen over hwcontext_vulkan.h).
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    crossbuild-essential-arm64 \
+    libavcodec-dev:arm64 libavformat-dev:arm64 libavutil-dev:arm64 libswscale-dev:arm64 \
+    libavfilter-dev:arm64 libavdevice-dev:arm64 \
+    libpipewire-0.3-dev:arm64 libopus-dev:arm64 \
+    libsdl3-dev:arm64 libgtk-4-dev:arm64 libadwaita-1-dev:arm64 \
+    libwayland-dev:arm64 libxkbcommon-dev:arm64 libvulkan-dev:arm64 \
+    && rm -rf /var/lib/apt/lists/*
+
+# 3. The Rust target — installed against the toolchain the WORKSPACE pins, not the image's
+#    default. The base image bakes whatever `stable` was at its build time, while every build
+#    in the repo switches to the exact channel in rust-toolchain.toml; adding the target to
+#    the default toolchain instead leaves the pinned one without an aarch64 std, and the build
+#    dies on `can't find crate for core` a few hundred crates in. Running rustup from a
+#    directory that contains the pin file resolves the right toolchain (and pre-downloads it,
+#    which every workspace job would otherwise pay for on first use).
+COPY rust-toolchain.toml /opt/pf-toolchain/
+WORKDIR /opt/pf-toolchain
+RUN rustup target add aarch64-unknown-linux-gnu && rustup show
+WORKDIR /
+
+# 4. Cross wiring. Everything in this image is a cross build, so the plain (un-suffixed)
+#    variables are safe and cover the crates that roll their own pkg-config/bindgen calls
+#    instead of going through the target-scoped lookups.
+#      * PKG_CONFIG uses Debian's multiarch wrapper, which resolves the arm64 .pc files and
+#        rewrites -I/-L into the sysroot without per-crate cooperation.
+#      * BINDGEN_EXTRA_CLANG_ARGS: clang defaults to the host triple, so bindgen would parse
+#        arm64 headers with amd64 type layouts (silently wrong, not a build error) — the
+#        explicit --target plus the multiarch include dir is what keeps the layouts honest.
+#      * CC_x86_64_unknown_linux_gnu routes HOST-targeted compiles through a wrapper that
+#        strips the arm64 include dirs — see ci/pf-host-cc for the ffmpeg-sys-next probe it
+#        exists for.
+COPY ci/pf-host-cc /usr/local/bin/pf-host-cc
+RUN chmod 0755 /usr/local/bin/pf-host-cc
+
+ENV CARGO_TARGET_AARCH64_UNKNOWN_LINUX_GNU_LINKER=aarch64-linux-gnu-gcc \
+    CC_aarch64_unknown_linux_gnu=aarch64-linux-gnu-gcc \
+    CXX_aarch64_unknown_linux_gnu=aarch64-linux-gnu-g++ \
+    AR_aarch64_unknown_linux_gnu=aarch64-linux-gnu-ar \
+    CC_x86_64_unknown_linux_gnu=/usr/local/bin/pf-host-cc \
+    PKG_CONFIG=aarch64-linux-gnu-pkg-config \
+    PKG_CONFIG_ALLOW_CROSS=1 \
+    BINDGEN_EXTRA_CLANG_ARGS="--target=aarch64-unknown-linux-gnu -I/usr/include/aarch64-linux-gnu"
+
+# Fail the BUILD, not some later CI job, if the wrapper or a sysroot .pc is missing.
+RUN command -v aarch64-linux-gnu-pkg-config \
+    && aarch64-linux-gnu-pkg-config --cflags libavcodec sdl3 gtk4 libpipewire-0.3 \
+    && aarch64-linux-gnu-gcc -dumpmachine | grep -q aarch64
@@ -3,7 +3,7 @@
 //! scenes.

 use crate::app::AppModel;
-use crate::trust::{KnownHost, KnownHosts};
+use crate::trust::{forget_placeholder, KnownHost, KnownHosts};
 use crate::ui_hosts::{ConnectRequest, HostsMsg};
 use gtk::glib;
 use gtk::prelude::*;
@@ -252,18 +252,6 @@ fn probe_all(hosts: &[KnownHost]) -> Vec<bool> {
    )
 }

-/// Drop an fp-less placeholder for `addr:port` (see `headless_pair`). No-op when none exists.
-fn forget_placeholder(addr: &str, port: u16) {
-    let mut known = KnownHosts::load();
-    let before = known.hosts.len();
-    known
-        .hosts
-        .retain(|h| !(h.fp_hex.is_empty() && h.addr == addr && h.port == port));
-    if known.hosts.len() != before {
-        let _ = known.save();
-    }
-}
-
 /// `--list-hosts [--probe]` — the saved known-hosts store as JSON (the store the Decky plugin
 /// renders). With `--probe`, each host carries an `online` bool from a live reachability probe
 /// (mDNS-independent); without it, `online` is `null` (unknown — the caller falls back to its
@@ -8,6 +8,7 @@ presenter of the Linux client re-architecture (punktfunk-planning:
 ```
 punktfunk-session --connect host[:port] [--fp HEX] [--launch id] [--fullscreen] [--stats]
 punktfunk-session --browse host[:port] [--mgmt PORT] [--fullscreen]
+punktfunk-session --pair <PIN> --connect host[:port] [--name LABEL]
 ```

 `--browse` opens the console game library (the Skia coverflow over the animated aurora)
@@ -17,9 +18,16 @@ pairing is the desktop client / Decky plugin's job. `PUNKTFUNK_FAKE_LIBRARY=<fil
 feeds canned entries with no host (portrait paths starting with `/` load from disk).

 Reads the same identity / known-hosts / settings stores as the desktop client
-(`punktfunk-client`) — pair there (or via its headless `--pair`) first; this binary never
+(`punktfunk-client`), so enrolling on either side makes the other work; this binary never
 connects to a host it has no pinned fingerprint for (`--fp HEX` overrides the store).

+`--pair <PIN> --connect host[:port]` runs the SPAKE2 ceremony with no window and no
+toolkit, prints `paired <addr>:<port> fp=<hex>`, and exits — the route for a machine that
+has only SSH (an embedded/kiosk client, an image being provisioned). `--name` sets the
+label the host files this client under, defaulting to the hostname. It is in the
+`--no-default-features` build too: enrolling must never be the reason a minimal image has
+to pull in Skia.
+
 Stdout is the machine interface: `{"ready":true}` after the first presented frame,
 `stats: …` once per second while the overlay tier isn't Off (always the full detailed
 text, whatever the OSD shows; `--stats` forces the overlay on), one
@@ -105,7 +105,7 @@ pub fn run(target: Option<&str>) -> u8 {
    };

    let opts = ConsoleOptions {
-        device_name: device_name(),
+        device_name: trust::device_name(),
        deck: is_steam_deck(),
    };
    let (overlay, handles) = match SkiaOverlay::console(opts, entry) {
@@ -251,22 +251,6 @@ pub fn run(target: Option<&str>) -> u8 {
    }
 }

-/// The machine's name — what the host lists this client as after pairing.
-fn device_name() -> String {
-    #[cfg(target_os = "linux")]
-    if let Ok(s) = std::fs::read_to_string("/etc/hostname") {
-        let s = s.trim();
-        if !s.is_empty() {
-            return s.to_string();
-        }
-    }
-    std::env::var("COMPUTERNAME")
-        .or_else(|_| std::env::var("HOSTNAME"))
-        .ok()
-        .filter(|s| !s.trim().is_empty())
-        .unwrap_or_else(|| "This device".into())
-}
-
 fn host_display_name(name: &str, addr: &str) -> String {
    if name.trim().is_empty() {
        addr.to_string()
@@ -5,8 +5,9 @@
 //! One stream session per invocation: `--connect host[:port]` (+ `--fp HEX`,
 //! `--launch id`, `--fullscreen`), exits when the session ends. Reads the same identity
 //! / known-hosts / settings stores as the desktop shell on each OS — the GTK client
-//! (`punktfunk-client`) on Linux, the WinUI client on Windows — so pairing there (or
-//! via the shell's headless `--pair`) makes this binary connect silently.
+//! (`punktfunk-client`) on Linux, the WinUI client on Windows — so pairing on either side
+//! makes the other connect silently. `--pair <PIN> --connect host` runs the ceremony here,
+//! with no window and no toolkit, for machines that have only a shell.
 //!
 //! Stdout is the machine interface (the shell↔session contract): `{"ready":true}` after
 //! the first presented frame, `stats:` lines per 1 s window, one `{"error": …}` /
@@ -61,6 +62,54 @@ mod session_main {
        Some((x.trim().parse().ok()?, y.trim().parse().ok()?))
    }

+    /// `--pair <PIN> --connect host[:port]` — the SPAKE2 PIN ceremony with no window, no GTK
+    /// and no console UI, so a machine that has only SSH can be enrolled: an embedded/kiosk
+    /// client, a headless box, an image being provisioned. Writes the verified host into the
+    /// same known-hosts store `--connect` reads, so pairing here is exactly what makes the
+    /// later stream connect silently.
+    ///
+    /// Deliberately identical in shape and output to `punktfunk-client --pair` (which stays
+    /// the desktop route) — the difference is only that this binary carries no toolkit, so it
+    /// is the one a minimal image installs. Present in the `--no-default-features` build too:
+    /// enrolment must not be the reason an embedded image has to pull in Skia.
+    fn headless_pair(pin: &str) -> u8 {
+        let Some(target) = arg_value("--connect") else {
+            eprintln!("--pair requires --connect host[:port]");
+            return EXIT_CONNECT_FAILED;
+        };
+        let (addr, port) = parse_host_port(&target);
+        // The label the HOST files this client under. A headless box has nobody to ask, so
+        // the hostname is the only name that will mean anything in the paired-devices list.
+        let name = arg_value("--name").unwrap_or_else(trust::device_name);
+
+        let identity = match trust::load_or_create_identity() {
+            Ok(i) => i,
+            Err(e) => {
+                eprintln!("client identity: {e:#}");
+                return EXIT_CONNECT_FAILED;
+            }
+        };
+        match trust::pair_with_host(&addr, port, &identity, pin, &name) {
+            Ok(fp) => {
+                let fp_hex = trust::hex(&fp);
+                trust::persist_host(
+                    &arg_value("--host-label").unwrap_or_else(|| addr.clone()),
+                    &addr,
+                    port,
+                    &fp_hex,
+                    true,
+                );
+                trust::forget_placeholder(&addr, port);
+                println!("paired {addr}:{port} fp={fp_hex}");
+                0
+            }
+            Err(e) => {
+                eprintln!("pairing failed: {} ({e:?})", trust::pair_error_message(&e));
+                EXIT_TRUST_REJECTED
+            }
+        }
+    }
+
    /// `host[:port]`, port defaulting to the native 9777.
    pub(crate) fn parse_host_port(target: &str) -> (String, u16) {
        match target.rsplit_once(':') {
@@ -312,6 +361,13 @@ mod session_main {
            };
        }

+        // `--pair <PIN>`: enrol this machine against a host and exit. Sits with the other
+        // non-streaming subcommands, above every graphics call — the box doing this may have
+        // no display at all.
+        if let Some(pin) = arg_value("--pair") {
+            return headless_pair(&pin);
+        }
+
        // Before any Vulkan call: make RADV expose its video-decode queue + extensions so the
        // decoder's `auto` path prefers Vulkan Video over VAAPI (Steam Deck, and any gated RADV).
        // Windows drivers (NVIDIA/AMD Adrenalin) expose theirs unconditionally.
@@ -369,12 +425,14 @@ mod session_main {
            eprintln!(
                "usage: punktfunk-session --connect host[:port] [--fp HEX] [--launch id] [--fullscreen]\n\
                 \x20      punktfunk-session --browse [host[:port]] [--mgmt PORT] [--fullscreen] [--json-status]\n\
+                 \x20      punktfunk-session --pair <PIN> --connect host[:port] [--name LABEL]\n\
                 \n\
                 Streams from a paired punktfunk host in a Vulkan window. --browse opens the\n\
                 gamepad console instead: bare --browse is the host list (discovery, PIN\n\
                 pairing, settings, wake-on-LAN); with a target it opens that host's game\n\
                 library. --connect never dials a host it has no pinned fingerprint for —\n\
-                 pair in the console or via `punktfunk-client --pair <PIN> --connect …`."
+                 enrol with --pair (no display needed), in the console, or from the desktop\n\
+                 client."
            );
            return EXIT_CONNECT_FAILED;
        };
@@ -406,7 +464,7 @@ mod session_main {
                "error",
                &format!(
                    "no pinned fingerprint for {addr}:{port} — pair first \
-                     (punktfunk-client --pair <PIN> --connect {addr}:{port}) or pass --fp HEX"
+                     (punktfunk-session --pair <PIN> --connect {addr}:{port}) or pass --fp HEX"
                ),
                Some(true),
            );
@@ -2191,7 +2191,15 @@ mod pipewire {
        // consumer imports raw dmabufs itself — the VAAPI backend (libva import + GPU CSC) or a
        // PyroWave session (the wavelet encoder's own Vulkan device, any vendor) → hand the raw
        // dmabuf straight to the encoder.
-        let vaapi_passthrough = zerocopy && !force_shm && importer.is_none() && raw_passthrough;
+        //
+        // ...unless the encoder already proved it cannot import them here. A driver that refuses
+        // the compositor's buffers refuses them identically on every retry, so without this the
+        // session died on its first frame and every reconnect repeated it. The latch (set by the
+        // encode side after consecutive import failures) is what turns that into one bad session
+        // followed by a working, if slower, host.
+        let raw_dmabuf_off = raw_passthrough && pf_zerocopy::raw_dmabuf_import_disabled();
+        let vaapi_passthrough =
+            zerocopy && !force_shm && importer.is_none() && raw_passthrough && !raw_dmabuf_off;
        // Producer-side NV12 (default-on; PUNKTFUNK_PIPEWIRE_NV12=0 escapes): gamescope offers a
        // one-fd LINEAR NV12 image when the consumer asks — its compositor pass does the RGB→YUV,
        // and the Vulkan Video encoder imports the buffer as its encode source directly (no host
@@ -2249,6 +2257,12 @@ mod pipewire {
            tracing::info!(
                "capture: PUNKTFUNK_FORCE_SHM — race-free SHM download path (no dmabuf, no zero-copy)"
            );
+        } else if raw_dmabuf_off {
+            tracing::warn!(
+                "zero-copy raw-dmabuf passthrough disabled after repeated encoder import failures \
+                 — capturing CPU frames instead (this host's GPU driver would not import the \
+                 compositor's buffers)"
+            );
        } else if zerocopy && !want_dmabuf {
            tracing::warn!("zero-copy: no importable dmabuf modifiers — using CPU path");
        } else if vaapi_passthrough && policy.pyrowave_modifiers.is_empty() {
@@ -2652,6 +2666,18 @@ mod pipewire {
            build_default_format_obj(preferred)
        };

+        // gamescope trap — the Steam overlay's presence in the stream is decided HERE by omission:
+        // gamescope's `paint_pipewire()` composites the overlay (Shift+Tab / Quick Access Menu) into
+        // the node it hands us ONLY when the consumer-negotiated `gamescope_focus_appid` is 0 — the
+        // default, and the "mirror the focused window + overlay" branch (gamescope ≥ 3.16.23; see
+        // `MIN_GAMESCOPE_OVERLAY`). None of the EnumFormat pods below advertise the
+        // `SPA_FORMAT_VIDEO_gamescope_focus_appid` property, so gamescope reads 0 and paints the
+        // overlay for us for free. DO NOT add a non-zero focus-appid (e.g. to "isolate the game" in a
+        // dedicated session) — that flips gamescope into the Remote-Play branch that deliberately
+        // drops the overlay (and all host chrome) back out of the capture. The cursor, external
+        // overlay (MangoHUD), and notifications are excluded from the node on EVERY gamescope
+        // version and are composited host-side instead (see `xfixes_cursor.rs`).
+        //
        // When zero-copy is on, offer ONLY a BGRx dmabuf format with our EGL-importable modifiers
        // (offering shm too makes the compositor pick shm). The modifier list is advertised with
        // DONT_FIXATE so the compositor's allocator chooses one; we re-emit the fixated format in
@@ -212,6 +212,40 @@ pub fn persist_host(name: &str, addr: &str, port: u16, fp_hex: &str, paired: boo
    let _ = known.save();
 }

+/// This machine's name — the label a host files this client under in its paired-devices list.
+/// `/etc/hostname` first (the answer on any Linux box, and the only one available in a minimal
+/// build with no GTK to ask), then the usual environment fallbacks.
+pub fn device_name() -> String {
+    #[cfg(target_os = "linux")]
+    if let Ok(s) = std::fs::read_to_string("/etc/hostname") {
+        let s = s.trim();
+        if !s.is_empty() {
+            return s.to_string();
+        }
+    }
+    std::env::var("COMPUTERNAME")
+        .or_else(|_| std::env::var("HOSTNAME"))
+        .ok()
+        .filter(|s| !s.trim().is_empty())
+        .unwrap_or_else(|| "This device".into())
+}
+
+/// Drop an fp-less placeholder entry for `addr:port`. A host added by address before any
+/// ceremony (`--add-host` with no `--fp`) is stored keyed by address with an empty fingerprint;
+/// once pairing yields the real one, [`persist_host`] writes a second, fp-keyed entry — so the
+/// placeholder has to go or the host list shows the same box twice. No-op (and no disk write)
+/// when there is none, which is the usual case.
+pub fn forget_placeholder(addr: &str, port: u16) {
+    let mut known = KnownHosts::load();
+    let before = known.hosts.len();
+    known
+        .hosts
+        .retain(|h| !(h.fp_hex.is_empty() && h.addr == addr && h.port == port));
+    if known.hosts.len() != before {
+        let _ = known.save();
+    }
+}
+
 /// Learn/refresh a saved host's Wake-on-LAN MAC(s) from its live advert (called while the host
 /// is online, matched by fingerprint or address). No-op — and no disk write — when unchanged, so
 /// the hosts page can call it on every discovery tick without churning the store.
@@ -6,11 +6,12 @@
 //! queue under the device's [`QueueLock`], fence-waited (sub-ms — Phase-0 measured
 //! 0.067 ms GPU at 1080p on the RTX 5070 Ti).
 //!
-//! Output: three separate R8 planes (Y full-res, Cb/Cr half-res) — the decode path
-//! requires STORAGE usage and IDENTITY/R swizzles, so the encoder's two-component
-//! RG8 trick is not allowed here (pyrowave.h validation). The presenter samples them
-//! with its planar CSC variant (BT.709 limited — the codec's fixed colour contract,
-//! there is no VUI). A small ring of plane-sets keeps a decode from overwriting the set
+//! Output: three separate single-component planes (Y full-res; Cb/Cr half-res, or
+//! full-res on a 4:4:4 session) — R8, or R16 UNORM on a 10-bit session — the decode
+//! path requires STORAGE usage and IDENTITY/R swizzles, so the encoder's two-component
+//! RG trick is not allowed here (pyrowave.h validation). The presenter samples them
+//! with its planar CSC variant (colour per the negotiated `ColorInfo` — the wavelet
+//! bitstream carries no VUI). A small ring of plane-sets keeps a decode from overwriting the set
 //! the presenter is still sampling; the synchronous fence bounds decode-side reuse and
 //! the ring depth covers present-side latency (≤ 1–2 frames in this pipeline).
 //!
@@ -222,9 +223,10 @@ unsafe extern "C" fn queue_unlock_cb(ud: *mut c_void) {
    unsafe { (*(ud as *const crate::video::QueueLock)).unlock() }
 }

-/// One decoded PyroWave frame: three R8 plane images on the presenter's device, GENERAL
-/// layout, decode-complete (the decoder fence-waits before handing it over). `slot`
-/// identifies the ring entry; the images/views live as long as the decoder.
+/// One decoded PyroWave frame: three single-component plane images (R8, or R16 on a
+/// 10-bit session) on the presenter's device, GENERAL layout, decode-complete (the
+/// decoder fence-waits before handing it over). `slot` identifies the ring entry; the
+/// images/views live as long as the decoder.
 pub struct PyroWavePlanarFrame {
    /// Raw `VkImageView`s (Y, Cb, Cr) for the presenter's planar CSC sampling.
    pub views: [u64; 3],
@@ -252,7 +254,8 @@ struct RetiredRing {
    retired_at: Instant,
 }

-/// One decode-output plane: R8, storage (decode writes) + sampled (presenter CSC).
+/// One decode-output plane (`fmt` = R8, or R16 on a 10-bit session): storage (decode
+/// writes) + sampled (presenter CSC).
 unsafe fn make_plane(
    device: &ash::Device,
    mem_props: &vk::PhysicalDeviceMemoryProperties,
@@ -539,7 +542,8 @@ impl PyroWaveDecoder {
            return Err(e);
        }

-        // Plane-set ring: 3 × R8, storage (decode writes) + sampled (presenter CSC).
+        // Plane-set ring: 3 single-component planes (R8; R16 on a 10-bit session),
+        // storage (decode writes) + sampled (presenter CSC).
        let mem_props = instance.get_physical_device_memory_properties(
            vk::PhysicalDevice::from_raw(vkd.physical_device as u64),
        );
@@ -873,12 +877,26 @@ impl PyroWaveDecoder {
            &vk::DependencyInfo::default().image_memory_barriers(&pre),
        );

+        // The declared format/extent MUST equal the ring image's real ones: pyrowave wraps
+        // our VkImage under `image_format` and vkCreateImageView's its storage view with
+        // `view_format` (pyrowave_c.cpp `WrappedViewBuffers::wrap` — its own
+        // `pyrowave_image_get_image_view` helper fills both from the image itself).
+        // Declaring R8 over a 10-bit session's R16_UNORM planes is an invalid view the
+        // driver executes anyway: its addressing covers half the surface, so decoded 8-bit
+        // codes fuse pairwise into 16-bit texels (structured garbage) and the never-written
+        // remainder samples as all-plane zeros (saturated green) — the 2026-07 AMD-client
+        // field report. Same discipline for chroma extents: a 4:4:4 ring is full-res.
+        let fmt = if self.hdr16 {
+            pw::VkFormat_VK_FORMAT_R16_UNORM
+        } else {
+            pw::VkFormat_VK_FORMAT_R8_UNORM
+        };
        let plane = |img: vk::Image, w: u32, h: u32| pw::pyrowave_image_view {
            image: img.as_raw() as usize as pw::VkImage,
            width: w,
            height: h,
-            image_format: pw::VkFormat_VK_FORMAT_R8_UNORM,
-            view_format: pw::VkFormat_VK_FORMAT_R8_UNORM,
+            image_format: fmt,
+            view_format: fmt,
            mip_level: 0,
            layer: 0,
            aspect: pw::VkImageAspectFlagBits_VK_IMAGE_ASPECT_COLOR_BIT,
@@ -886,11 +904,16 @@ impl PyroWaveDecoder {
            layout: pw::VkImageLayout_VK_IMAGE_LAYOUT_GENERAL,
        };
        let (w, h) = (self.width, self.height);
+        let (cw, ch) = if self.chroma444 {
+            (w, h)
+        } else {
+            (w / 2, h / 2)
+        };
        let buffers = pw::pyrowave_gpu_buffers {
            planes: [
                plane(self.ring[slot].imgs[0], w, h),
-                plane(self.ring[slot].imgs[1], w / 2, h / 2),
-                plane(self.ring[slot].imgs[2], w / 2, h / 2),
+                plane(self.ring[slot].imgs[1], cw, ch),
+                plane(self.ring[slot].imgs[2], cw, ch),
            ],
        };
        pw::pyrowave_device_set_command_buffer(
@@ -7,6 +7,45 @@
 use anyhow::Result;
 use pf_frame::CapturedFrame;

+/// Whether an encoder fed `format` must be built 10-bit — decided by **the pixels that actually
+/// arrive**, never by the negotiated `bit_depth`.
+///
+/// The three Windows backends each derived this as `bit_depth >= 10 || matches!(format, P010 |
+/// Rgb10a2)`, i.e. the *negotiated* depth could force a 10-bit encoder over an 8-bit capture. That
+/// combination is not hypothetical: a client advertises 10-bit, the handshake negotiates
+/// `bit_depth = 10`, and then enabling advanced colour on the IDD virtual display fails — at which
+/// point the capturer says so and delivers 8-bit NV12 anyway (`pf-capture`'s idd_push logs "10-bit
+/// HDR was negotiated but enabling advanced color on the virtual display FAILED — encoding 8-bit
+/// SDR"). Every backend then lost the session, each in its own way: native AMF and native QSV
+/// `bail!` at open because the format does not match the P010 they derived, and the libavcodec
+/// path accepted the open and then failed EVERY submit forever (its per-frame depth check
+/// recomputes from the frame, which never matches), where `reset()` could not help because the
+/// rebuild re-derived the same wrong depth.
+///
+/// Following the pixels keeps the stream alive and, more importantly, keeps it HONEST: the depth
+/// also selects the colour signalling (BT.2020 PQ vs BT.709) and the staging surface format, so an
+/// 8-bit capture now yields an 8-bit stream that says it is SDR — which is what the capturer
+/// already reported it is sending. The negotiated depth remains an upper bound; the session label
+/// may still claim HDR, and that mismatch belongs to the negotiation, not to the encoder.
+/// Windows-only: the three backends that derive an encoder depth from a capture live there
+/// (native AMF, native QSV, libavcodec AMF/QSV). The Linux backends take the depth from the
+/// negotiated `bit_depth` alone because their capture formats carry it unambiguously.
+#[cfg(target_os = "windows")]
+pub(crate) fn ten_bit_input(format: pf_frame::PixelFormat, negotiated_depth: u8) -> bool {
+    use pf_frame::PixelFormat;
+    let ten = matches!(format, PixelFormat::P010 | PixelFormat::Rgb10a2);
+    if negotiated_depth >= 10 && !ten {
+        tracing::warn!(
+            ?format,
+            negotiated_depth,
+            "session negotiated 10-bit but the capturer delivers an 8-bit format — encoding 8-bit \
+             SDR (the stream's colour signalling follows the pixels; check whether advanced colour \
+             failed to enable on the virtual display)"
+        );
+    }
+    ten
+}
+
 /// An encoded access unit (one NAL/AU) to hand to `punktfunk_core` for FEC + packetization.
 /// `data` is in-band Annex-B (the encoder is opened without a global header), so each
 /// keyframe carries its own VPS/SPS/PPS — the bytes are both a playable elementary
@@ -73,7 +112,7 @@ impl AuChunk {
 }

 /// Codec selection negotiated with the client.
-#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
 pub enum Codec {
    H264,
    H265,
@@ -255,6 +294,19 @@ pub struct EncoderCaps {
    /// `USER_FLAG_RECOVERY_POINT` on every Nth emitted AU, re-phased at each IDR). 0 when intra-refresh
    /// is off. Only consulted when [`intra_refresh_recovery`](Self::intra_refresh_recovery) is set.
    pub intra_refresh_period: u32,
+    /// The encoder composites [`CapturedFrame::cursor`] into the picture it encodes.
+    ///
+    /// `open_video`'s `cursor_blend` argument is a REQUEST, and for most of this crate's life it was
+    /// nothing else: `lib.rs` literally did `let _ = cursor_blend;` and only three backends ever read
+    /// `frame.cursor`. So a session could ask for a composited pointer, get a backend that silently
+    /// discards it, and stream with no mouse cursor at all — the confirmed symptom on the VAAPI
+    /// dmabuf path and the libav-NVENC CUDA path.
+    ///
+    /// This makes the answer queryable instead of assumed. It is deliberately a plain fact about the
+    /// encoder, not a policy: what to DO when a session wants blending and the backend cannot is the
+    /// host's call, since only the host can re-plan capture (fall back to capturer-side compositing).
+    /// `open_video` can only warn, which it does.
+    pub blends_cursor: bool,
 }

 /// A hardware encoder. One per session; runs on the encode thread.
@@ -313,8 +365,15 @@ pub trait Encoder: Send {
    /// encoder analog of the capturer depth escalation: AUs ride ~one loop tick behind (`poll`
    /// may return `None` while an encode is in flight) in exchange for capture/submit no longer
    /// serializing on the encode wait. Returns whether pipelined retrieve is (now) active; the
-    /// switch may be deferred to the next safe point internally. `false` from the default impl =
-    /// unsupported — the session loop stops asking. De-escalation is a v2 item everywhere.
+    /// switch may be deferred to the next safe point internally. `set_pipelined(true)` returning
+    /// `false` (the default impl) = unsupported — the session loop stops asking.
+    ///
+    /// `set_pipelined(false)` requests the wind-back (de-escalation, latency recovery): the
+    /// backend restores its sync-retrieve mode — and the latency features that mode carries
+    /// (IO-stream binding, sub-frame chunking) — at its next safe point, usually via a session
+    /// rebuild whose first frame is an IDR. The return is still "is pipelined retrieve active":
+    /// the caller polls until it reads `false`. Backends that never escalate return `false`
+    /// trivially. An operator pin (`PUNKTFUNK_NVENC_ASYNC=1`) refuses the wind-back.
    fn set_pipelined(&mut self, _on: bool) -> bool {
        false
    }
@@ -360,6 +419,16 @@ pub trait Encoder: Send {
    fn reconfigure_bitrate(&mut self, _bps: u64) -> bool {
        false
    }
+    /// The bitrate (bps) the encoder is ACTUALLY running at (or will open at, for a lazily-opened
+    /// backend) — the encoder-side truth after any internal clamp, e.g. the direct-NVENC
+    /// codec-level ceiling search. The session loop reads this after every open/reconfigure and
+    /// stores IT, not the requested rate, as the live bitrate — so the send pacer, the console
+    /// and the client controller's ack all track what the ASIC really targets (a controller fed
+    /// the requested rate keeps climbing from a phantom base, §ABR overdrive). `None` (the
+    /// default) = the backend doesn't track an applied rate; the caller keeps the requested one.
+    fn applied_bitrate_bps(&self) -> Option<u64> {
+        None
+    }
    /// Wire-chunk the encoder's AUs at the session's shard payload size (the PyroWave
    /// datagram-aligned mode, plan §4.4): every `shard_payload` window of the emitted AU
    /// starts a fresh self-delimiting codec packet, zero-padded to the window — so a lost
@@ -379,6 +448,20 @@ pub trait Encoder: Send {
    fn set_input_ring_depth(&mut self, _depth: usize) {}
    /// Signal end-of-stream. After this, drain the remaining AUs with [`poll`](Self::poll)
    /// until it returns `None` — NVENC buffers frames internally even at `delay=0`.
+    ///
+    /// **The two production encode loops deliberately do not call this**, and that is not an
+    /// oversight to be "fixed" by a later sweep. Both reach their exit only after the transport is
+    /// already gone (the client disconnected, or the session was stopped), so the AUs a flush would
+    /// recover have nowhere to go — while flushing is the one call on this trait that can BLOCK on a
+    /// wedged encoder, on precisely the teardown path a stopped session needs to complete promptly.
+    /// The Linux direct-SDK NVENC backend makes that concrete: its retrieve-thread join is untimed
+    /// (see the note in `enc/linux/nvenc_cuda.rs`), so a flush there could hang a session that is
+    /// already ending.
+    ///
+    /// It is kept rather than deleted because it does have real consumers: the `spike` dev
+    /// subcommand, which encodes a FINITE clip and genuinely wants the tail, and the `#[ignore]`d
+    /// hardware smoke tests across the backends, which assert the drain contract on real GPUs.
+    /// Those are finite-stream users; a live session is not one.
    fn flush(&mut self) -> Result<()>;
 }

@@ -413,6 +496,17 @@ impl Codec {
    }
 }

+/// Pixel rate (luma samples/s) at or above which NVENC split-frame encoding is FORCED 2-way —
+/// one number shared by the direct-SDK selector (`nvenc_core::resolve_split_mode`) and the libav
+/// `split_encode_mode` option author (`linux::NvencEncoder`), so the two paths can never disagree
+/// about which modes split. A single NVENC engine tops out ~1 Gpix/s on HEVC, and AUTO doesn't
+/// engage below ~2112 px height, so the sessions that need the second engine must be forced. Set
+/// BELOW 1 Gpix/s deliberately: 4K120 — the mode this threshold exists for — is 3840×2160×120 =
+/// 995,328,000, which a `> 1_000_000_000` gate missed by 0.47% and left on AUTO (pinned ~107 fps
+/// on a 4090). 950 M keeps margin for fractional refresh rates while leaving 1440p240 (884.7 M,
+/// comfortably single-engine) on AUTO.
+pub const SPLIT_FORCE_PIXEL_RATE: u64 = 950_000_000;
+
 /// `PUNKTFUNK_VBV_FRAMES` — HRD/VBV size in frame intervals (default 1.0, the strict low-latency
 /// shape every backend ships: each frame must fit its rate share, keeping frame sizes uniform for
 /// the pacer). The AMF/VAAPI/QSV paths parse the same variable locally; this helper brings the
@@ -426,6 +520,35 @@ pub(crate) fn vbv_frames_env() -> f64 {
        .unwrap_or(1.0)
 }

+/// The same HRD/VBV window as [`vbv_frames_env`], expressed the way the Vulkan Video encode API
+/// wants it: `(virtualBufferSizeInMs, initialVirtualBufferSizeInMs)`.
+///
+/// Every other backend states the window in **bits** (`bitrate / fps × frames`); Vulkan states it
+/// in **milliseconds**. `vulkan_video.rs` consumes this ONLY when the driver advertises VBR
+/// (WP6.3): a tight window under CBR makes the driver stuff underspent frames with filler NALs up
+/// to the exact rate share — measured 97 % filler on the 780M — because CBR must keep the CPB from
+/// overflowing and Vulkan exposes no filler-suppression control. VBR permits the underspend, so
+/// the tight window only ever *bounds* a complex frame.
+///
+/// The initial fill stays at half the window, preserving the RATIO the hardcoded (1000, 500)
+/// pair had — the direct-NVENC house shape uses a FULL-window initial fill instead; measured on
+/// RADV the difference is inert (the firmware showed no window sensitivity at all). Both
+/// VUIDs on `VkVideoEncodeRateControlInfoKHR`'s window fields are satisfied by construction: the
+/// window clamps to `>= 1` so it is non-zero, and `window / 2 <= window` always
+/// (`VUID-...-08358` is `<=`, relaxed in Vulkan 1.3.299).
+///
+/// Carries its only caller's gate: `vulkan_video.rs` is the sole ms-form consumer, and with the
+/// crate-wide `allow(dead_code)` gone (WP0.3) an item unused in ANY feature combination is a hard
+/// error — this is dead on every Windows leg.
+#[cfg(all(target_os = "linux", feature = "vulkan-encode"))]
+pub(crate) fn vbv_window_ms(fps: u32) -> (u32, u32) {
+    let frames = vbv_frames_env();
+    let ms = (frames * 1000.0 / fps.max(1) as f64).round();
+    // `f64 as u32` saturates at the bounds in Rust, so an absurd `PUNKTFUNK_VBV_FRAMES` cannot wrap.
+    let window = (ms as u32).max(1);
+    (window, window / 2)
+}
+
 /// Validate a requested encode resolution before we allocate buffers or open NVENC. Rejects
 /// zero/odd-sized and out-of-range modes with a clear error instead of letting buffer math
 /// overflow or the encoder open fail with an opaque NVENC code. A client can request any
@@ -454,6 +577,26 @@ pub fn validate_dimensions(codec: Codec, width: u32, height: u32) -> Result<()>
             (use HEVC/AV1 above 4096, or lower the client resolution)"
        );
    }
+    // PyroWave's vendored rate controller packs the 32×32 block index into the low 16 bits of
+    // `RDOperation::block_offset_saving` (pyrowave-sys `patches/0002-rdo-saving-clamp.patch`).
+    // Past `u16::MAX` blocks the index collides with the `saving` field, the resolve over-credits,
+    // and the emitted payload can overshoot the buffer `pyrowave_encoder_packetize` writes into —
+    // whose only bounds check is an `assert` that the Release (NDEBUG) vendored build compiles out.
+    // So this is a hard cap, not a quality knob.
+    //
+    // Checked against 4:2:0, the *most permissive* chroma: a mode that cannot fit even there can
+    // fit no PyroWave session at all, so it belongs at this single chokepoint (which both the
+    // negotiator and `open_video_backend` run) rather than only in the per-backend opens. 4:4:4
+    // has twice the block count and is checked again at open, where the real chroma is known —
+    // and the negotiator's 4:4:4 → 4:2:0 downgrade means an oversized mode arrives at the encoder
+    // as 4:2:0, which is exactly the case the old open-time guard skipped.
+    #[cfg(feature = "pyrowave")]
+    if codec == Codec::PyroWave && !crate::pyrowave_mode_fits_rdo(width, height, false) {
+        anyhow::bail!(
+            "invalid PyroWave resolution {width}x{height}: exceeds the rate controller's 16-bit \
+             block index (pyrowave-sys patches/0002) — lower the client resolution"
+        );
+    }
    Ok(())
 }

@@ -461,6 +604,31 @@ pub fn validate_dimensions(codec: Codec, width: u32, height: u32) -> Result<()>
 mod tests {
    use super::*;

+    /// WP6.3. The window VUIDs on `VkVideoEncodeRateControlInfoKHR` are the whole contract of
+    /// this helper, and both are edge cases: the window must be non-zero (a high-refresh mode
+    /// rounds a sub-1 ms window down to nothing) and the initial fill must be at most the window
+    /// (`<=` — `VUID-...-08358` was relaxed in 1.3.299). Env-free so it pins the default shape —
+    /// the scaled cases belong to whoever sets `PUNKTFUNK_VBV_FRAMES`. Carries the helper's own
+    /// cfg gate (see its note), so it runs on the Linux `vulkan-encode` leg.
+    #[cfg(all(target_os = "linux", feature = "vulkan-encode"))]
+    #[test]
+    fn vbv_window_is_about_one_frame_and_always_legal() {
+        // The house default is ~1 frame interval, not the 1000 ms the Vulkan backend hardwired.
+        assert_eq!(vbv_window_ms(60).0, 17); // 16.67 ms
+        assert_eq!(vbv_window_ms(30).0, 33);
+        assert_eq!(vbv_window_ms(240).0, 4);
+        for fps in [1, 24, 30, 60, 120, 144, 240, 480, 1000, 4000, u32::MAX] {
+            let (window, initial) = vbv_window_ms(fps);
+            assert!(window > 0, "virtualBufferSizeInMs must be > 0 (fps {fps})");
+            assert!(
+                initial <= window,
+                "initialVirtualBufferSizeInMs must be <= virtualBufferSizeInMs (fps {fps})"
+            );
+        }
+        // fps 0 must not divide by zero — `open` clamps, but the helper is called directly too.
+        assert!(vbv_window_ms(0).0 > 0);
+    }
+
    #[test]
    fn rejects_zero_and_odd_dimensions() {
        assert!(validate_dimensions(Codec::H265, 0, 1080).is_err());
@@ -477,6 +645,26 @@ mod tests {
        assert!(validate_dimensions(Codec::H264, 3840, 4098).is_err());
    }

+    /// PyroWave's hard cap is the rate controller's 16-bit block index, not just
+    /// `max_dimension()`. Checked at 4:2:0 (the most permissive chroma), because a mode that
+    /// cannot fit there cannot fit at any chroma — and because the negotiator's 4:4:4 → 4:2:0
+    /// downgrade delivers oversized modes to the encoder AS 4:2:0. HEVC/AV1 at the same
+    /// dimensions must stay unaffected.
+    #[cfg(feature = "pyrowave")]
+    #[test]
+    fn pyrowave_rejects_modes_past_the_rdo_block_index() {
+        // Fits: 8K 4:2:0 is 49125 blocks.
+        assert!(validate_dimensions(Codec::PyroWave, 7680, 4320).is_ok());
+        // Does not fit at 4:2:0 (73728 / 98304 blocks) — must be refused even though both are
+        // within `Codec::PyroWave.max_dimension()` (8192).
+        assert!(validate_dimensions(Codec::PyroWave, 8192, 6144).is_err());
+        assert!(validate_dimensions(Codec::PyroWave, 8192, 8192).is_err());
+        // The same modes remain legal for the H.26x/AV1 codecs, which have no such rate
+        // controller — the cap must not leak across codecs.
+        assert!(validate_dimensions(Codec::H265, 8192, 8192).is_ok());
+        assert!(validate_dimensions(Codec::Av1, 8192, 6144).is_ok());
+    }
+
    #[test]
    fn hevc_and_av1_allow_up_to_8192() {
        for c in [Codec::H265, Codec::Av1] {
@@ -180,7 +180,6 @@ pub struct NvencEncoder {
    /// This session opened as full-chroma 4:4:4 (FREXT) — via either input path.
    want_444: bool,
    src_format: PixelFormat,
-    expand: bool,
    width: u32,
    height: u32,
    fps: u32,
@@ -413,57 +412,6 @@ impl NvencEncoder {
            None
        };

-        // CPU CSC paths: build the packed-RGB → planar swscale (no rescale) into the encoder's
-        // input frame. Two users: 4:4:4 (RGB→YUV444P, BT.709, range per the flag) and HDR
-        // (X2RGB10/X2BGR10→P010, BT.2020 limited — the PQ transfer is per-channel and rides
-        // through the matrix untouched). Skipped on the zero-copy path (`cuda`): the worker's GPU
-        // convert already delivers ready CUDA frames — no CPU pixels exist to scale.
-        let sws_csc = if (want_444 || want_hdr10) && !cuda {
-            let src_av = pixel_to_av(sws_src_pixel(format)?);
-            let dst_av = pixel_to_av(nvenc_pixel);
-            // SAFETY: `sws_getContext` allocates a swscale context for the given src/dst dims + pixel
-            // formats. Both dims are the encoder's positive `width`/`height` as `c_int`; `src_av` is a
-            // valid `AVPixelFormat` (from the `sws_src_pixel`-validated packed-RGB source), the dst is
-            // YUV444P (4:4:4) or P010LE (HDR). The trailing filter/param pointers are null = "use
-            // defaults" (documented as accepted). No Rust memory is borrowed; the returned pointer is
-            // null-checked below.
-            let sws = unsafe {
-                ffi::sws_getContext(
-                    width as c_int,
-                    height as c_int,
-                    src_av,
-                    width as c_int,
-                    height as c_int,
-                    dst_av,
-                    SWS_POINT,
-                    ptr::null_mut(),
-                    ptr::null_mut(),
-                    ptr::null(),
-                )
-            };
-            if sws.is_null() {
-                bail!("sws_getContext(RGB→{nvenc_pixel:?}) failed");
-            }
-            // SAFETY: `sws` is the non-null context from the call above (null-checked). The
-            // coefficient tables from `sws_getCoefficients` (ITU-709 for 4:4:4, BT.2020 NCL for HDR
-            // — matching the VUI written above) are process-lifetime libswscale statics, reused for
-            // src+dst matrices; `sws_setColorspaceDetails` only reads them and writes scalar CSC
-            // settings into `sws` (dstRange matches the VUI: 0 = limited, 1 = the
-            // PUNKTFUNK_444_FULLRANGE experiment; HDR is always limited). No Rust memory is passed.
-            unsafe {
-                let cs = ffi::sws_getCoefficients(if want_hdr10 {
-                    super::libav::SWS_CS_BT2020
-                } else {
-                    SWS_CS_ITU709
-                });
-                let dst_range = i32::from(full_range_444);
-                ffi::sws_setColorspaceDetails(sws, cs, 1, cs, dst_range, 0, 1 << 16, 1 << 16);
-            }
-            Some(sws)
-        } else {
-            None
-        };
-
        // Low-latency NVENC tuning (plan §7 / linux-setup doc).
        let mut opts = Dictionary::new();
        opts.set("preset", "p1"); // fastest
@@ -491,15 +439,19 @@ impl NvencEncoder {
        }

        // Split-frame encode across both NVENC engines (GB203 has 2) when the pixel rate exceeds
-        // a single engine's HEVC capacity (~1 Gpix/s); e.g. 5120x1440@240 = 1.77 Gpix/s needs it,
-        // @120 = 0.88 Gpix/s does not. HEVC/AV1 only (not H.264). AUTO won't engage below ~2112px
+        // a single engine's HEVC capacity; e.g. 5120x1440@240 = 1.77 Gpix/s needs it, @120
+        // (0.88 Gpix/s) does not. HEVC/AV1 only (not H.264). AUTO won't engage below ~2112px
        // height, so we force `2`; below the threshold we leave it AUTO (split costs ~2% BD-rate).
-        // Output is standard HEVC — transparent to the client. Override with PUNKTFUNK_SPLIT_ENCODE.
+        // Threshold shared with the direct-SDK selector ([`super::SPLIT_FORCE_PIXEL_RATE`] — set
+        // so 4K120 = 995.3 Mpix/s forces, which `> 1e9` famously missed by 0.47%). Output is
+        // standard HEVC — transparent to the client. Override with PUNKTFUNK_SPLIT_ENCODE.
        let pix_rate = width as u64 * height as u64 * fps as u64;
        let split = std::env::var("PUNKTFUNK_SPLIT_ENCODE").ok();
        match split.as_deref() {
            Some(mode) => opts.set("split_encode_mode", mode),
-            None if matches!(codec, Codec::H265 | Codec::Av1) && pix_rate > 1_000_000_000 => {
+            None if matches!(codec, Codec::H265 | Codec::Av1)
+                && pix_rate >= super::SPLIT_FORCE_PIXEL_RATE =>
+            {
                opts.set("split_encode_mode", "2");
                tracing::info!(
                    pix_rate,
@@ -549,6 +501,84 @@ impl NvencEncoder {
            );
        }

+        // Built HERE, below the fallible encoder open, NOT above it. `sws_getContext` returns a raw
+        // pointer whose only free is `Drop for NvencEncoder` — and `Drop` needs a CONSTRUCTED
+        // `Self`, which does not exist on `open`'s early returns (the intra-refresh-unsupported
+        // retry, which recurses into `Self::open`, and the plain error return). Creating the
+        // context above them leaked one per failed attempt, and `open_nvenc_probed`'s EINVAL
+        // bitrate ladder calls `open` up to ~10 times, so a host stepping its bitrate down leaked a
+        // context per step. Nothing between here and the `Ok(NvencEncoder { … })` below can return,
+        // so this placement makes the leak unrepresentable rather than merely unlikely.
+        // CPU CSC paths: build the packed-RGB → planar swscale (no rescale) into the encoder's
+        // input frame. THREE users: 4:4:4 (RGB→YUV444P, BT.709, range per the flag), HDR
+        // (X2RGB10/X2BGR10→P010, BT.2020 limited — the PQ transfer is per-channel and rides
+        // through the matrix untouched), and the packed 3-bpp expand (RGB24/BGR24→rgb0/bgr0).
+        //
+        // The expand used to be a hand-written per-pixel loop in `submit_cpu`: `w*h` iterations,
+        // each building two bounds-checked sub-slices for a 3-byte copy — a shape LLVM will not
+        // vectorise into the byte shuffle it is, on the COMMON CPU path (the portal and wlroots
+        // both commonly fixate packed 24-bit RGB, and pf-capture offers it first). swscale's
+        // packed-RGB expanders are SIMD, the sibling VAAPI backend already routes RGB24/BGR24
+        // through them, and this file already owned the context lifecycle — so the change is net
+        // subtractive. The three are mutually exclusive by construction: `expand` is only ever
+        // true on the packed-RGB 4:2:0 path (see `nvenc_pixel`/`expand` above), never with
+        // `want_444`, so one context serves whichever applies.
+        //
+        // Skipped on the zero-copy path (`cuda`): the worker's GPU convert already delivers ready
+        // CUDA frames — no CPU pixels exist to scale.
+        let sws_csc = if (want_444 || want_hdr10 || expand) && !cuda {
+            let src_av = pixel_to_av(sws_src_pixel(format)?);
+            let dst_av = pixel_to_av(nvenc_pixel);
+            // SAFETY: `sws_getContext` allocates a swscale context for the given src/dst dims + pixel
+            // formats. Both dims are the encoder's positive `width`/`height` as `c_int`; `src_av` is a
+            // valid `AVPixelFormat` (from the `sws_src_pixel`-validated packed-RGB source), the dst is
+            // YUV444P (4:4:4) or P010LE (HDR). The trailing filter/param pointers are null = "use
+            // defaults" (documented as accepted). No Rust memory is borrowed; the returned pointer is
+            // null-checked below.
+            let sws = unsafe {
+                ffi::sws_getContext(
+                    width as c_int,
+                    height as c_int,
+                    src_av,
+                    width as c_int,
+                    height as c_int,
+                    dst_av,
+                    SWS_POINT,
+                    ptr::null_mut(),
+                    ptr::null_mut(),
+                    ptr::null(),
+                )
+            };
+            if sws.is_null() {
+                bail!("sws_getContext(RGB→{nvenc_pixel:?}) failed");
+            }
+            // Colour math applies to the CSC users ONLY. The expand is a pure byte shuffle —
+            // packed 3-bpp RGB/BGR to the same channels in 4 bytes, `nvenc_pixel` being `rgb0`/
+            // `bgr0` — and NVENC does the RGB→YUV itself downstream. Handing it a matrix + range
+            // here would silently range-convert every packed-RGB session, which is exactly what the
+            // module header promises does not happen ("no colour math").
+            if want_444 || want_hdr10 {
+                // SAFETY: `sws` is the non-null context from the call above (null-checked). The
+                // coefficient tables from `sws_getCoefficients` (ITU-709 for 4:4:4, BT.2020 NCL for
+                // HDR — matching the VUI written above) are process-lifetime libswscale statics,
+                // reused for src+dst matrices; `sws_setColorspaceDetails` only reads them and writes
+                // scalar CSC settings into `sws` (dstRange matches the VUI: 0 = limited, 1 = the
+                // PUNKTFUNK_444_FULLRANGE experiment; HDR is always limited). No Rust memory is passed.
+                unsafe {
+                    let cs = ffi::sws_getCoefficients(if want_hdr10 {
+                        super::libav::SWS_CS_BT2020
+                    } else {
+                        SWS_CS_ITU709
+                    });
+                    let dst_range = i32::from(full_range_444);
+                    ffi::sws_setColorspaceDetails(sws, cs, 1, cs, dst_range, 0, 1 << 16, 1 << 16);
+                }
+            }
+            Some(sws)
+        } else {
+            None
+        };
+
        let frame = if cuda {
            None
        } else {
@@ -561,7 +591,6 @@ impl NvencEncoder {
            sws_csc,
            want_444,
            src_format: format,
-            expand,
            width,
            height,
            fps,
@@ -591,6 +620,9 @@ impl NvencEncoder {
 impl Encoder for NvencEncoder {
    fn caps(&self) -> super::EncoderCaps {
        super::EncoderCaps {
+            // libav NVENC hands the frame straight to the encoder — `frame.cursor` is never read,
+            // so a cursor-as-metadata session loses its pointer on this backend (audit finding).
+            blends_cursor: false,
            // 4:4:4 iff this session opened FREXT — the CPU swscale path or the zero-copy GPU
            // convert. RFI/HDR-SEI stay unsupported on libavcodec NVENC (the trait defaults).
            chroma_444: self.want_444,
@@ -695,8 +727,10 @@ impl NvencEncoder {
            bytes.len(),
            src_row * h
        );
-        // 4:4:4: swscale the packed RGB straight into the planar YUV444P input frame (BT.709 limited),
-        // then send it — no byte-expand. The 4:2:0 RGB path (below) feeds NVENC packed RGB directly.
+        // swscale the packed RGB straight into the encoder's input frame, then send it. Serves all
+        // three CSC users (see `open`): 4:4:4 → planar YUV444P, HDR → P010, and the packed 3-bpp
+        // expand → `rgb0`/`bgr0`. The remaining branch below is the 4-bpp source, which needs no
+        // conversion at all — just a row copy honouring the destination stride.
        if let Some(sws) = self.sws_csc {
            let frame = self
                .frame
@@ -705,10 +739,13 @@ impl NvencEncoder {
            // SAFETY: `format == self.src_format` and `bytes.len() >= src_row * h` (the `ensure!`s
            // above), so `sws_scale` reads `h` rows of `src_row` bytes from `src_data[0] = bytes`
            // (packed RGB is single-plane; the other src planes are null/0) — all in bounds. `sws` is
-            // the non-null context built in `open`. The dst is `frame`'s underlying `AVFrame`: its
-            // `data`/`linesize` in-struct arrays were sized for YUV444P by `VideoFrame::new`, and the
-            // 3 planes are each `width`×`height`. All pointers are live locals for this synchronous
-            // call; the encoder runs only on this thread (`unsafe impl Send`), so no aliasing/race.
+            // the non-null context built in `open`. The dst is `frame`'s underlying `AVFrame`, whose
+            // `data`/`linesize` in-struct arrays were sized by `VideoFrame::new` for the very
+            // `nvenc_pixel` this context was built to output — 3 planes of `width`×`height` for
+            // YUV444P, 2 for P010, 1 packed plane for `rgb0`/`bgr0` — so swscale writes exactly the
+            // planes it allocated, at the strides it reports. All pointers are live locals for this
+            // synchronous call; the encoder runs only on this thread (`unsafe impl Send`), so no
+            // aliasing/race.
            unsafe {
                let dst_av = frame.as_mut_ptr();
                let src_data: [*const u8; 4] =
@@ -724,7 +761,7 @@ impl NvencEncoder {
                    (*dst_av).linesize.as_ptr(),
                );
                if r < 0 {
-                    bail!("sws_scale(RGB→YUV444P) failed ({r})");
+                    bail!("sws_scale(CPU CSC → encoder input) failed ({r})");
                }
            }
            frame.set_pts(Some(pts));
@@ -733,7 +770,7 @@ impl NvencEncoder {
            } else {
                ffmpeg::picture::Type::None
            });
-            self.enc.send_frame(frame).context("send_frame(444)")?;
+            self.enc.send_frame(frame).context("send_frame(swscale)")?;
            return Ok(());
        }
        let frame = self
@@ -742,18 +779,9 @@ impl NvencEncoder {
            .context("CPU frame missing (encoder opened in CUDA mode)")?;
        let stride = frame.stride(0); // dst is 4-bpp, aligned
        let dst = frame.data_mut(0);
-        if self.expand {
-            // packed 3-bpp RGB/BGR → 4-bpp *0 (copy 3 bytes, zero the pad byte)
-            for y in 0..h {
-                let s = &bytes[y * src_row..y * src_row + src_row];
-                let drow = &mut dst[y * stride..y * stride + w * 4];
-                for x in 0..w {
-                    drow[x * 4..x * 4 + 3].copy_from_slice(&s[x * 3..x * 3 + 3]);
-                    drow[x * 4 + 3] = 0;
-                }
-            }
-        } else {
-            // 4-bpp → 4-bpp, honoring the (possibly larger) dst stride
+        {
+            // 4-bpp → 4-bpp, honoring the (possibly larger) dst stride. The 3-bpp expand that used
+            // to live here as a per-pixel loop is now swscale's job (see the branch above).
            for y in 0..h {
                dst[y * stride..y * stride + src_row]
                    .copy_from_slice(&bytes[y * src_row..y * src_row + src_row]);
@@ -876,6 +904,58 @@ impl Drop for NvencEncoder {
    }
 }

+/// Serialises the save → `AV_LOG_FATAL` → restore window that every capability probe opens around
+/// an encoder open it *expects* to fail.
+///
+/// libav's log level is one process-global `int`, and the probes race each other for real: the
+/// NVENC and VAAPI 4:4:4/10-bit probes are reached from `/serverinfo` and from session bring-up.
+/// Two overlapping save/restore pairs interleave as get(INFO) → get(FATAL) → set(INFO) →
+/// set(FATAL), and the process is then pinned at `AV_LOG_FATAL` for good — every later libav
+/// diagnostic silently dropped, which is precisely the logging you want when a stream later fails
+/// to open. The probes run process-once and already cost a real encoder open, so serialising them
+/// costs nothing measurable.
+static LIBAV_LOG_LEVEL: std::sync::Mutex<()> = std::sync::Mutex::new(());
+
+/// RAII quiet-window over libav's global log level: drops it to `AV_LOG_FATAL` on construction and
+/// restores the previous level on drop, holding [`LIBAV_LOG_LEVEL`] for the whole window.
+///
+/// Callers must have completed `ffmpeg::init()` first. Not re-entrant — no probe may construct a
+/// second guard while holding one (none do; the probe bodies only reach encoder-open helpers).
+/// `pub(crate)` so the VAAPI probes share the one lock: they race the NVENC probes on the same
+/// global.
+pub(crate) struct QuietLibavLog {
+    prev: c_int,
+    // Held for the lifetime of the guard. `Drop for QuietLibavLog` runs before the struct's fields
+    // are dropped, so the restore below still happens under the lock.
+    _lock: std::sync::MutexGuard<'static, ()>,
+}
+
+impl QuietLibavLog {
+    pub(crate) fn new() -> Self {
+        // Poison-tolerant: a probe that panicked mid-window already restored the level via `Drop`,
+        // and refusing the lock forever afterwards would be a worse outcome than proceeding.
+        let lock = LIBAV_LOG_LEVEL
+            .lock()
+            .unwrap_or_else(|poisoned| poisoned.into_inner());
+        // SAFETY: libav is initialized by the caller; `av_log_{get,set}_level` only read/write the
+        // global int level (no pointer args) and are always sound post-init.
+        let prev = unsafe {
+            let p = ffi::av_log_get_level();
+            ffi::av_log_set_level(ffi::AV_LOG_FATAL);
+            p
+        };
+        Self { prev, _lock: lock }
+    }
+}
+
+impl Drop for QuietLibavLog {
+    fn drop(&mut self) {
+        // SAFETY: restore the saved global level (scalar arg, no pointers); libav was initialized
+        // before this guard was constructed.
+        unsafe { ffi::av_log_set_level(self.prev) };
+    }
+}
+
 /// Probe whether this NVIDIA GPU + driver + libavcodec can actually encode HEVC **4:4:4** (Range
 /// Extensions). Opens a tiny real `hevc_nvenc` 4:4:4 session — the exact path [`NvencEncoder::open`]
 /// takes for a live 4:4:4 stream — and reports whether it succeeded. HEVC-only; the result is cached
@@ -889,14 +969,9 @@ pub fn probe_can_encode_444(codec: Codec) -> bool {
        return false;
    }
    // Quiet ffmpeg's open error on a GPU that lacks 4:4:4 — the probe failing is an expected outcome.
-    // SAFETY: libav initialized above; `av_log_{get,set}_level` only read/write the global int level
-    // (no pointer args) and are always sound post-init.
-    let prev = unsafe {
-        let p = ffi::av_log_get_level();
-        ffi::av_log_set_level(ffi::AV_LOG_FATAL);
-        p
-    };
-    let ok = NvencEncoder::open(
+    // Held until the function returns, so the level is restored after the open either way.
+    let _quiet = QuietLibavLog::new();
+    NvencEncoder::open(
        codec,
        PixelFormat::Bgra,
        640,
@@ -907,10 +982,7 @@ pub fn probe_can_encode_444(codec: Codec) -> bool {
        8,
        ChromaFormat::Yuv444,
    )
-    .is_ok();
-    // SAFETY: restore the saved global log level (scalar arg, no pointers).
-    unsafe { ffi::av_log_set_level(prev) };
-    ok
+    .is_ok()
 }

 /// Probe whether this NVIDIA GPU + driver + libavcodec can actually encode 10-bit (HEVC Main10 /
@@ -926,14 +998,9 @@ pub fn probe_can_encode_10bit(codec: Codec) -> bool {
        return false;
    }
    // Quiet ffmpeg's open error on a GPU that lacks 10-bit — the probe failing is an expected outcome.
-    // SAFETY: libav initialized above; `av_log_{get,set}_level` only read/write the global int level
-    // (no pointer args) and are always sound post-init.
-    let prev = unsafe {
-        let p = ffi::av_log_get_level();
-        ffi::av_log_set_level(ffi::AV_LOG_FATAL);
-        p
-    };
-    let ok = NvencEncoder::open(
+    // Held until the function returns, so the level is restored after the open either way.
+    let _quiet = QuietLibavLog::new();
+    NvencEncoder::open(
        codec,
        PixelFormat::X2Rgb10,
        640,
@@ -944,10 +1011,7 @@ pub fn probe_can_encode_10bit(codec: Codec) -> bool {
        10,
        ChromaFormat::Yuv420,
    )
-    .is_ok();
-    // SAFETY: restore the saved global log level (scalar arg, no pointers).
-    unsafe { ffi::av_log_set_level(prev) };
-    ok
+    .is_ok()
 }

 #[cfg(test)]
@@ -61,8 +61,9 @@
 #![deny(clippy::undocumented_unsafe_blocks)]

 use super::nvenc_core::{
-    apply_low_latency_config, build_init_params, codec_guid, resolve_slices, resolve_subframe,
-    LowLatencyConfig, NvStatusExt, RFI_DPB,
+    apply_low_latency_config, build_init_params, cached_ceiling, codec_guid, plan_range_recovery,
+    resolve_slices, resolve_split_mode, resolve_subframe, store_ceiling, CeilingKey,
+    LowLatencyConfig, NvStatusExt, RangePlan,
 };
 use super::nvenc_status;
 use super::{AuChunk, ChromaFormat, Codec, EncodedFrame, Encoder, EncoderCaps};
@@ -268,13 +269,19 @@ fn async_retrieve_requested() -> bool {

 /// Max encodes in flight in two-thread mode (`PUNKTFUNK_NVENC_ASYNC_DEPTH`, default 4, clamped
 /// `2..=POOL-1` — a bitstream must never be reused mid-encode, and the input ring is the same
-/// depth). Mirrors the Windows knob exactly.
+/// depth). Mirrors the Windows knob exactly, memoization included: this is the backpressure
+/// **loop condition** in `submit`, so an engaged two-thread session re-read the environment once
+/// per spin. The default session never pays it (the condition short-circuits on `async_rt`), which
+/// is why the audit's severity ranking for this site was inverted — but an escalated one did.
 fn async_inflight_cap() -> usize {
-    std::env::var("PUNKTFUNK_NVENC_ASYNC_DEPTH")
-        .ok()
-        .and_then(|s| s.parse::<usize>().ok())
-        .unwrap_or(4)
-        .clamp(2, POOL - 1)
+    static CAP: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
+    *CAP.get_or_init(|| {
+        std::env::var("PUNKTFUNK_NVENC_ASYNC_DEPTH")
+            .ok()
+            .and_then(|s| s.trim().parse::<usize>().ok())
+            .unwrap_or(4)
+            .clamp(2, POOL - 1)
+    })
 }

 /// Stream-ordered submit (`PUNKTFUNK_NVENC_STREAM_ORDERED`, default ON; `0` = the pre-existing
@@ -595,6 +602,12 @@ pub struct NvencCudaEncoder {
    /// (escalate-and-hold, like the depth escalation); the switch itself happens at the next
    /// safe point via [`maybe_engage_async`](Self::maybe_engage_async).
    want_async: bool,
+    /// A de-escalation request ([`Encoder::set_pipelined(false)`]) waiting for its safe point:
+    /// the next drained moment tears the session down and lazily re-inits SYNC (IO-stream
+    /// binding and sub-frame chunking re-arm at that re-init). Distinct from `!want_async` —
+    /// an operator-forced async session (`PUNKTFUNK_NVENC_ASYNC=1`) also has `want_async`
+    /// false, and a de-escalation must never tear THAT down.
+    want_sync: bool,
    /// Boxed `CUstream` the session's IO-stream binding points at (`NvEncSetIOCudaStreams` takes
    /// POINTERS to `CUstream`, and this struct moves — the pointee needs a stable heap address for
    /// the session's lifetime). Null when stream-ordering is off; freed in `teardown` AFTER the
@@ -702,6 +715,7 @@ impl NvencCudaEncoder {
            last_rfi_range: None,
            async_rt: None,
            want_async: false,
+            want_sync: false,
            io_stream: ptr::null_mut(),
            stream_ordered: false,
            slices: 1,
@@ -735,6 +749,29 @@ impl NvencCudaEncoder {
        }
    }

+    /// [`maybe_engage_async`](Self::maybe_engage_async)'s inverse — wind the escalated pipelined
+    /// retrieve back at a safe point: nothing in flight, then a clean session rebuild whose lazy
+    /// SYNC re-init restores the IO-stream binding and re-arms sub-frame chunking (the two
+    /// latency features the escalation traded away). No-op until
+    /// [`want_sync`](Self::want_sync) is set and `pending` drains.
+    fn maybe_disengage_async(&mut self) {
+        if !self.want_sync || self.async_rt.is_none() || !self.pending.is_empty() {
+            return;
+        }
+        self.want_sync = false;
+        if self.inited {
+            // SAFETY: encode thread, `pending` empty ⇒ no encode in flight (and nothing queued
+            // to the retrieve thread); `teardown` joins the retrieve thread and handles exactly
+            // this live-session state — the next submit lazily re-inits sync.
+            unsafe { self.teardown() };
+            tracing::info!(
+                "NVENC pipelined-retrieve de-escalation: rebuilding the session with the sync \
+                 retrieve (IO-stream binding and sub-frame chunking restored); next frame opens \
+                 with an IDR"
+            );
+        }
+    }
+
    /// Tear down the encode session + pooled resources. Reused on a size change and at Drop.
    unsafe fn teardown(&mut self) {
        if self.encoder.is_null() {
@@ -839,6 +876,9 @@ impl NvencCudaEncoder {
                e,
            ));
        }
+        // The handshake with the kernel module just succeeded — from here on, an
+        // `NV_ENC_ERR_INVALID_VERSION` in this process cannot be a driver version skew.
+        nvenc_status::note_session_opened();
        let wmax = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_WIDTH_MAX);
        let hmax = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_HEIGHT_MAX);
        let yuv444 = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_YUV444_ENCODE);
@@ -1000,6 +1040,23 @@ impl NvencCudaEncoder {
        Ok(cfg)
    }

+    /// This session config's identity in the process-lifetime bitrate-ceiling cache
+    /// (`nvenc_core::{cached_ceiling, store_ceiling}`). GPU identity is the process-global shared
+    /// `CUcontext` pointer — one context per process, stable for its lifetime; only valid once
+    /// `cu_ctx` is bound (`init_session` start), which every caller is downstream of.
+    fn ceiling_key(&self, split_mode: u32) -> CeilingKey {
+        CeilingKey {
+            gpu: self.cu_ctx as u64,
+            codec: self.codec,
+            width: self.width,
+            height: self.height,
+            fps: self.fps,
+            bit_depth: self.bit_depth,
+            chroma_444: self.chroma_444,
+            split_mode,
+        }
+    }
+
    /// Open + configure + initialize ONE NVENC CUDA session at `bitrate` (bps) and `split_mode`.
    /// Returns the session handle, or destroys it and returns the error.
    unsafe fn try_open_session(&self, bitrate: u64, split_mode: u32) -> Result<*mut c_void> {
@@ -1019,6 +1076,7 @@ impl NvencCudaEncoder {
            }
            return Err(nvenc_status::call_err("open_encode_session_ex", e));
        }
+        nvenc_status::note_session_opened();

        let mut cfg = match self.build_config(enc, bitrate) {
            Ok(cfg) => cfg,
@@ -1074,58 +1132,71 @@ impl NvencCudaEncoder {
            }
            const FLOOR_BPS: u64 = 10_000_000;
            let requested_bps = self.bitrate_bps;
-            // 2-way NVENC split-frame encoding (Ada dual-NVENC) above ~1 Gpix/s; env override
-            // PUNKTFUNK_SPLIT_ENCODE = 0/disable | 1/auto | 2 | 3. HEVC/AV1 only.
+            // 2-way NVENC split-frame encoding (Ada dual-NVENC) — shared selector, see
+            // [`resolve_split_mode`] for the precedence (env override / 10-bit / pixel rate).
            let pixel_rate = self.width as u64 * self.height as u64 * self.fps.max(1) as u64;
-            let mut split_mode: u32 = match std::env::var("PUNKTFUNK_SPLIT_ENCODE").ok().as_deref()
-            {
-                Some("0") | Some("disable") => {
-                    nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32
-                }
-                Some("1") | Some("auto") => {
-                    nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_AUTO_FORCED_MODE as u32
-                }
-                Some("3") => nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_THREE_FORCED_MODE as u32,
-                Some("2") => nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_TWO_FORCED_MODE as u32,
-                _ if self.bit_depth >= 10 => {
-                    nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32
-                }
-                _ if pixel_rate > 1_000_000_000 => {
-                    nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_TWO_FORCED_MODE as u32
-                }
-                _ => nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_AUTO_MODE as u32,
-            };
+            let split_mode: u32 = resolve_split_mode(self.bit_depth, pixel_rate);
            const CLAMP_TOL_BPS: u64 = 20_000_000;

-            let mut probe = self.try_open_session(requested_bps, split_mode);
-            // Disambiguate a forced-split rejection from a bitrate-cap rejection.
+            // Ceiling cache (process lifetime, `nvenc_core`): a prior clamp search already found
+            // this config's max accepted rate — open straight AT the ceiling instead of paying
+            // the ~6-open binary search (and its session churn) on every ABR overshoot.
+            let mut target_bps = requested_bps;
+            if let Some(ceiling) = cached_ceiling(&self.ceiling_key(split_mode)) {
+                if requested_bps > ceiling {
+                    tracing::info!(
+                        requested_mbps = requested_bps / 1_000_000,
+                        ceiling_mbps = ceiling / 1_000_000,
+                        "NVENC (Linux): requested bitrate above the cached codec-level ceiling — \
+                         opening at the ceiling"
+                    );
+                    target_bps = ceiling;
+                }
+            }
+
+            let mut probe = self.try_open_session(target_bps, split_mode);
+            // The cache is advisory: a stale entry (driver change, identity collision) must not
+            // wedge the open — retry the requested rate and let the search below rediscover.
+            if probe.is_err() && target_bps < requested_bps {
+                target_bps = requested_bps;
+                probe = self.try_open_session(requested_bps, split_mode);
+            }
+            // Disambiguate a forced-split rejection from a bitrate-cap rejection. `used_split`
+            // tracks the mode sessions ACTUALLY open with from here on — it feeds
+            // `self.split_mode` (a reconfigure must re-present it) and the ceiling-cache key.
+            let mut used_split = split_mode;
            let split_on =
                split_mode != nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32;
            if probe.is_err() && split_on {
                let no_split = nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32;
-                if let Ok(e) = self.try_open_session(requested_bps, no_split) {
+                if let Ok(e) = self.try_open_session(target_bps, no_split) {
                    tracing::warn!(
                        "NVENC (Linux): split-encode rejected by codec/config — disabled"
                    );
-                    split_mode = no_split;
+                    used_split = no_split;
                    probe = Ok(e);
                }
            }

            let enc = match probe {
                Ok(enc) => {
-                    self.bitrate_bps = requested_bps;
+                    self.bitrate_bps = target_bps;
                    enc
                }
+                // Only a parameter/caps rejection means "the bitrate is above the codec-level
+                // ceiling". A transient failure (busy engine, session limit, OOM, device loss,
+                // version skew) must propagate — a search steered by it would discover, and
+                // cache, a bogus ceiling.
+                Err(e) if !nvenc_status::is_param_rejection(&e) => return Err(e),
                Err(_) => {
                    // Requested bitrate exceeds the codec-level ceiling — binary-search the max accepted.
                    let mut lo = FLOOR_BPS;
-                    let mut hi = requested_bps;
+                    let mut hi = target_bps;
                    let mut best: *mut c_void = ptr::null_mut();
                    let mut best_bps = 0u64;
                    while hi > lo + CLAMP_TOL_BPS {
                        let mid = lo + (hi - lo) / 2;
-                        match self.try_open_session(mid, split_mode) {
+                        match self.try_open_session(mid, used_split) {
                            Ok(e) => {
                                if !best.is_null() {
                                    let _ = (api().destroy_encoder)(best);
@@ -1134,18 +1205,30 @@ impl NvencCudaEncoder {
                                best_bps = mid;
                                lo = mid;
                            }
-                            Err(_) => hi = mid,
+                            Err(e) if nvenc_status::is_param_rejection(&e) => hi = mid,
+                            Err(e) => {
+                                // Environmental mid-search failure: don't let it shrink the
+                                // search — release the partial result and propagate.
+                                if !best.is_null() {
+                                    let _ = (api().destroy_encoder)(best);
+                                }
+                                return Err(e);
+                            }
                        }
                    }
                    if best.is_null() {
                        let no_split =
                            nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32;
-                        best = self
-                            .try_open_session(FLOOR_BPS, split_mode)
-                            .or_else(|_| self.try_open_session(FLOOR_BPS, no_split))
-                            .context(
-                                "NVENC initialize_encoder rejected even at the floor bitrate",
-                            )?;
+                        best = match self.try_open_session(FLOOR_BPS, used_split) {
+                            Ok(e) => e,
+                            Err(_) => {
+                                let e = self.try_open_session(FLOOR_BPS, no_split).context(
+                                    "NVENC initialize_encoder rejected even at the floor bitrate",
+                                )?;
+                                used_split = no_split;
+                                e
+                            }
+                        };
                        best_bps = FLOOR_BPS;
                    }
                    tracing::warn!(
@@ -1153,15 +1236,13 @@ impl NvencCudaEncoder {
                        clamped_mbps = best_bps / 1_000_000,
                        "NVENC (Linux): requested bitrate above the GPU codec-level ceiling — clamped"
                    );
+                    store_ceiling(self.ceiling_key(used_split), best_bps);
                    self.bitrate_bps = best_bps;
                    best
                }
            };
            self.encoder = enc;
-            // (Best effort: the floor fallback above may have succeeded split-disabled without
-            // updating `split_mode` — a later reconfigure then presents the forced mode, NVENC
-            // rejects it, and the caller's rebuild fallback covers the mismatch.)
-            self.split_mode = split_mode;
+            self.split_mode = used_split;

            // Output bitstream pool.
            for _ in 0..POOL {
@@ -1345,6 +1426,10 @@ impl NvencCudaEncoder {
                mbps = self.bitrate_bps / 1_000_000,
                codec = ?self.codec_guid,
                fmt = ?self.buffer_fmt,
+                // The FINAL split mode (post any rejection fallback) at INFO — journals run
+                // INFO+, and "did 4K120 actually split across engines?" was undiagnosable from
+                // a user log without it (Windows only had a debug! at selection time).
+                split_mode = self.split_mode,
                "NVENC CUDA session ready"
            );
            Ok(())
@@ -1426,9 +1511,10 @@ impl Encoder for NvencCudaEncoder {
                "Linux direct-NVENC needs a CUDA frame (FramePayload::Cuda); got a CPU/dmabuf frame"
            ),
        };
-        // A pending pipelined-retrieve escalation engages here, at the submit-side safe point
-        // (nothing in flight after the previous poll drained).
+        // A pending pipelined-retrieve escalation — or de-escalation — engages here, at the
+        // submit-side safe point (nothing in flight after the previous poll drained).
        self.maybe_engage_async();
+        self.maybe_disengage_async();
        // Re-init on a size change (the capturer can return at a different resolution after a mode
        // switch). Format changes (NV12↔YUV444) likewise re-init.
        let new_fmt = buffer_format(buf);
@@ -1560,25 +1646,6 @@ impl Encoder for NvencCudaEncoder {
                    }
                } else {
                    self.cursor_blend_warned = false;
-                    // TEMP KWin composite probe (DROP BEFORE MERGE): prove the blend dispatches
-                    // and with what geometry.
-                    {
-                        use std::sync::atomic::{AtomicU64, Ordering as ProbeOrd};
-                        static PROBE_BLEND: AtomicU64 = AtomicU64::new(0);
-                        let n = PROBE_BLEND.fetch_add(1, ProbeOrd::Relaxed) + 1;
-                        if n == 1 || n % 512 == 0 {
-                            tracing::info!(
-                                n,
-                                fmt = ?self.buffer_fmt,
-                                ov_x = ov.x,
-                                ov_y = ov.y,
-                                ov_w = ov.w,
-                                ov_h = ov.h,
-                                visible = ov.visible,
-                                "cursor blend probe: vulkan dispatch submitted"
-                            );
-                        }
-                    }
                }
            } else if !self.cursor_blend_warned {
                self.cursor_blend_warned = true;
@@ -1736,12 +1803,25 @@ impl Encoder for NvencCudaEncoder {

    fn set_pipelined(&mut self, on: bool) -> bool {
        if !on {
-            // v1 is escalate-and-hold (no de-escalation), mirroring the depth escalation.
+            // De-escalation (the v2 of escalate-and-hold): latch the wind-back intent; the
+            // switch itself happens at the next drained safe point
+            // ([`maybe_disengage_async`](Self::maybe_disengage_async)) — the caller polls
+            // this same method until it reports inactive.
+            if async_retrieve_env() == Some(true) {
+                // Operator pinned async on — de-escalation must not undo an explicit choice.
+                return self.want_async || self.async_rt.is_some();
+            }
+            if self.want_async || self.async_rt.is_some() {
+                self.want_async = false;
+                self.want_sync = true;
+                self.maybe_disengage_async();
+            }
            return self.want_async || self.async_rt.is_some();
        }
        if async_retrieve_env() == Some(false) {
            return false; // operator veto: PUNKTFUNK_NVENC_ASYNC=0 means NEVER
        }
+        self.want_sync = false; // latest intent wins — cancel a pending wind-back
        if !self.want_async && self.async_rt.is_none() {
            self.want_async = true;
            self.maybe_engage_async();
@@ -1751,6 +1831,8 @@ impl Encoder for NvencCudaEncoder {

    fn caps(&self) -> EncoderCaps {
        EncoderCaps {
+            // Composites `frame.cursor` via the SPIR-V blend over the Vulkan-allocated input slot.
+            blends_cursor: true,
            supports_rfi: self.rfi_supported,
            supports_hdr_metadata: self.hdr,
            chroma_444: self.chroma_444,
@@ -1765,48 +1847,46 @@ impl Encoder for NvencCudaEncoder {
    }

    fn invalidate_ref_frames(&mut self, first: i64, last: i64) -> bool {
-        if self.encoder.is_null() || !self.rfi_supported || first < 0 || first > last {
+        // Range validity, covering-range dedup, DPB window and clamp all live in
+        // `nvenc_core::plan_range_recovery` — one policy for both direct-NVENC backends; only the
+        // session gate and the driver loop are this backend's.
+        if self.encoder.is_null() || !self.rfi_supported {
            return false;
        }
-        // Already invalidated a covering range for this loss event — re-arm the anchor (the previous
-        // anchor AU may itself have been lost) but skip the driver calls.
-        if let Some((pf, pl)) = self.last_rfi_range {
-            if first >= pf && last <= pl {
+        match plan_range_recovery(first, last, self.frame_idx, self.last_rfi_range) {
+            // Already invalidated a covering range for this loss event — re-arm the anchor (the
+            // previous anchor AU may itself have been lost) but skip the driver calls.
+            RangePlan::Covered => {
                self.pending_anchor = true;
-                return true;
+                true
            }
-        }
-        // The DPB holds `[frame_idx - RFI_DPB, frame_idx - 1]`; a lost frame older than that can't be
-        // invalidated, so the only correct recovery is an IDR.
-        let oldest_in_dpb = self.frame_idx - RFI_DPB as i64;
-        if first < oldest_in_dpb {
-            return false;
-        }
-        let last = last.min(self.frame_idx - 1);
-        if first > last {
-            return false;
-        }
-        // Each input's `inputTimeStamp` is the WIRE frame index (pinned by `submit_indexed`), so the
-        // client's lost-frame range maps 1:1 onto the timestamps NVENC invalidates here.
-        // SAFETY: `invalidate_ref_frames` is a function pointer from the runtime table; `self.encoder`
-        // was checked non-null and is the live session; this runs on the encode thread (no concurrent
-        // NVENC use). Each `ts` is clamped to `[oldest_in_dpb, frame_idx - 1]`, naming a frame still
-        // in the DPB; the call passes only that `u64` (no struct).
-        unsafe {
-            for ts in first..=last {
-                if (api().invalidate_ref_frames)(self.encoder, ts as u64)
-                    .nv_ok()
-                    .is_err()
-                {
-                    return false;
+            RangePlan::Decline => false,
+            RangePlan::Invalidate { first, last } => {
+                // Each input's `inputTimeStamp` is the WIRE frame index (pinned by
+                // `submit_indexed`), so the client's lost-frame range maps 1:1 onto the timestamps
+                // NVENC invalidates here.
+                // SAFETY: `invalidate_ref_frames` is a function pointer from the runtime table;
+                // `self.encoder` was checked non-null and is the live session; this runs on the
+                // encode thread (no concurrent NVENC use). The plan clamped each `ts` to
+                // `[oldest_in_dpb, frame_idx - 1]`, naming a frame still in the DPB; the call
+                // passes only that `u64` (no struct).
+                unsafe {
+                    for ts in first..=last {
+                        if (api().invalidate_ref_frames)(self.encoder, ts as u64)
+                            .nv_ok()
+                            .is_err()
+                        {
+                            return false;
+                        }
+                    }
                }
+                self.last_rfi_range = Some((first, last));
+                // The next submitted frame is the clean re-anchor — arm the tag so its AU ships
+                // with `recovery_anchor` and the client lifts its post-loss freeze on it.
+                self.pending_anchor = true;
+                true
            }
        }
-        self.last_rfi_range = Some((first, last));
-        // The next submitted frame is the clean re-anchor — arm the tag so its AU ships with
-        // `recovery_anchor` and the client lifts its post-loss freeze on it.
-        self.pending_anchor = true;
-        true
    }

    fn poll(&mut self) -> Result<Option<EncodedFrame>> {
@@ -2061,6 +2141,15 @@ impl Encoder for NvencCudaEncoder {
            self.bitrate_bps = bps;
            return true;
        }
+        // Cached codec-level ceiling: clamp the target BEFORE the driver call, so a known
+        // overshoot retargets to the ceiling IN PLACE instead of bouncing off the driver into
+        // the caller's full-rebuild fallback (an IDR plus ~half a second of session churn per
+        // ABR overshoot on the pre-cache path). The caller reads the clamp back through
+        // [`Encoder::applied_bitrate_bps`].
+        let bps = match cached_ceiling(&self.ceiling_key(self.split_mode)) {
+            Some(ceiling) => bps.min(ceiling),
+            None => bps,
+        };
        // SAFETY: `inited` ⟹ `self.encoder` is the live session and every call here runs on the
        // encode thread with no NVENC call in flight (the session loop calls this between
        // submit/poll). `build_config` only queries the preset on that session; `cfg` outlives
@@ -2108,6 +2197,12 @@ impl Encoder for NvencCudaEncoder {
        }
    }

+    fn applied_bitrate_bps(&self) -> Option<u64> {
+        // `bitrate_bps` is the post-clamp truth: the open path's ceiling search and the
+        // reconfigure path's cache clamp both write what the session ACTUALLY targets.
+        Some(self.bitrate_bps)
+    }
+
    fn flush(&mut self) -> Result<()> {
        Ok(()) // P1/ULL + frameIntervalP=1: each submit yields its AU; no internal queue to drain.
    }
@@ -2165,6 +2260,7 @@ mod tests {
            true,
            8,
            ChromaFormat::Yuv420,
+            false,
        )
        .expect("open NVENC CUDA session");

@@ -2241,6 +2337,7 @@ mod tests {
            true,
            8,
            ChromaFormat::Yuv444,
+            false,
        )
        .expect("open NVENC CUDA 4:4:4 session");

@@ -2286,6 +2383,7 @@ mod tests {
            true,
            8,
            ChromaFormat::Yuv420,
+            false,
        )
        .expect("open NVENC CUDA session");

@@ -2343,6 +2441,7 @@ mod tests {
            true,
            8,
            ChromaFormat::Yuv420,
+            false,
        ) else {
            eprintln!(
                "skipping rfi_declines_impossible_ranges: NVENC unavailable (no NVIDIA driver)"
@@ -2369,6 +2468,7 @@ mod tests {
            true,
            8,
            ChromaFormat::Yuv420,
+            false,
        )
        .expect("open NVENC CUDA encoder")
    }
@@ -2403,6 +2503,7 @@ mod tests {
                true,
                8,
                ChromaFormat::Yuv420,
+                false,
            )
            .expect("open");
            for f in 0..4u32 {
@@ -2540,6 +2641,7 @@ mod tests {
            true,
            8,
            ChromaFormat::Yuv420,
+            false,
        )
        .expect("open NVENC CUDA session");
        let frame = nv12_frame(W, H, 0);
@@ -2581,6 +2683,7 @@ mod tests {
            true,
            8,
            ChromaFormat::Yuv420,
+            false,
        )
        .expect("open NVENC CUDA session");
        // Steady sync frames first (stream-ordered mode).
@@ -2660,6 +2763,7 @@ mod tests {
            true,
            8,
            ChromaFormat::Yuv420,
+            false,
        )
        .expect("open NVENC CUDA session");

@@ -2765,6 +2869,7 @@ mod tests {
            true,
            8,
            ChromaFormat::Yuv420,
+            false,
        )
        .expect("open NVENC CUDA session");

@@ -28,22 +28,18 @@ use ffmpeg::format::Pixel;
 use ffmpeg::{codec, encoder, Dictionary};
 use ffmpeg_next as ffmpeg;
 use pf_frame::{CapturedFrame, DmabufFrame, FramePayload, PixelFormat};
+use std::collections::HashMap;
 use std::ffi::{CStr, CString};
 use std::os::fd::AsRawFd;
 use std::os::raw::c_int;
 use std::ptr;
-use std::sync::atomic::{AtomicU8, Ordering};
+use std::sync::{Mutex, OnceLock};

 use super::libav::{
    apply_low_latency_rc, pixel_to_av, poll_encoder, PollOutcome, SWS_CS_ITU709, SWS_POINT,
 };
 use ffmpeg::ffi; // = ffmpeg_sys_next

-/// `fourcc(a,b,c,d)` — DRM FourCC packing (`a | b<<8 | c<<16 | d<<24`).
-const fn fourcc(a: u8, b: u8, c: u8, d: u8) -> u32 {
-    (a as u32) | ((b as u32) << 8) | ((c as u32) << 16) | ((d as u32) << 24)
-}
-
 /// The render node a VAAPI/DRM device should open, from [`pf_gpu::linux_render_node`]: a
 /// matched web-console GPU preference pins it, else `PUNKTFUNK_RENDER_NODE`, else the single-GPU
 /// default.
@@ -71,21 +67,38 @@ fn vaapi_sws_src(format: PixelFormat) -> Result<Pixel> {
    })
 }

-/// Which VAAPI entrypoint mode opened successfully, cached per codec (index = [`lp_idx`]):
-/// 0 = unknown, 1 = default (full-feature `EncSlice`), 2 = low-power (`EncSliceLP`/VDEnc).
-/// Modern Intel (Gen12+/Arc) removed the full-feature encode entrypoints, so the default open
-/// fails there and only `low_power=1` works; AMD (radeonsi) is the reverse. Caching the resolved
-/// mode lets later sessions/probes skip the known-failing attempt (and its libav error spew).
-static LP_MODE: [AtomicU8; 3] = [AtomicU8::new(0), AtomicU8::new(0), AtomicU8::new(0)];
+/// Which VAAPI entrypoint mode opened successfully: 1 = default (full-feature `EncSlice`),
+/// 2 = low-power (`EncSliceLP`/VDEnc). Modern Intel (Gen12+/Arc) removed the full-feature encode
+/// entrypoints, so the default open fails there and only `low_power=1` works; AMD (radeonsi) is the
+/// reverse. Caching the resolved mode lets later sessions/probes skip the known-failing attempt
+/// (and its libav error spew).
+///
+/// Keyed on **(render node, codec, bit depth)**, and every part of that key is load-bearing:
+///
+/// * The render node, because the entrypoint is a property of the DEVICE libva opens — and
+///   `render_node()` follows the web-console GPU preference. This used to be a process-global array
+///   keyed by codec alone, which made it a session-killer rather than a staleness bug: once a mode
+///   is latched the open tries exactly ONE mode and the `[false, true]` fallback is gone. Latch
+///   low-power on an Intel Arc, switch the preference to an AMD dGPU, and every VAAPI open there
+///   passes `low_power=1` — which radeonsi rejects — with no full-feature retry, for the process
+///   lifetime: the probe reports all-false AND the session's own encoder open fails.
+///   Keyed on the node rather than `pf_gpu::selection_key()` on purpose — the node is literally what
+///   `render_node()` hands libva, so key and device cannot describe different GPUs.
+/// * The bit depth, because Main10 and 8-bit can resolve to different entrypoints on the same
+///   device; one shared slot let an 8-bit answer pin the 10-bit open (HDR under-advertisement).
+static LP_MODE: OnceLock<Mutex<HashMap<LpKey, u8>>> = OnceLock::new();

-fn lp_idx(codec: Codec) -> usize {
-    match codec {
-        Codec::H264 => 0,
-        Codec::H265 => 1,
-        Codec::Av1 => 2,
-        // Guarded by the open_video dispatch: PyroWave never opens the VAAPI backend.
-        Codec::PyroWave => unreachable!("PyroWave has no VAAPI encoder"),
-    }
+/// (render-node path, codec label, 10-bit) — see [`LP_MODE`].
+type LpKey = (String, &'static str, bool);
+
+/// The [`LP_MODE`] key for this device/codec/depth. `render_node()` is re-read rather than cached
+/// so a GPU-preference change is picked up on the next open.
+fn lp_key(codec: Codec, ten_bit: bool) -> LpKey {
+    (
+        render_node().to_string_lossy().into_owned(),
+        codec.label(),
+        ten_bit,
+    )
 }

 /// `PUNKTFUNK_VAAPI_LOW_POWER` pins the entrypoint mode (`1` = low-power only, `0` = full-feature
@@ -116,11 +129,16 @@ unsafe fn open_vaapi_encoder(
    frames_ref: *mut ffi::AVBufferRef,
    ten_bit: bool,
 ) -> Result<encoder::video::Encoder> {
-    let idx = lp_idx(codec);
+    let key = lp_key(codec, ten_bit);
+    let cached = LP_MODE
+        .get_or_init(|| Mutex::new(HashMap::new()))
+        .lock()
+        .map(|m| m.get(&key).copied().unwrap_or(0))
+        .unwrap_or(0);
    let modes: &[bool] = match low_power_override() {
        Some(true) => &[true],
        Some(false) => &[false],
-        None => match LP_MODE[idx].load(Ordering::Relaxed) {
+        None => match cached {
            1 => &[false],
            2 => &[true],
            _ => &[false, true],
@@ -140,7 +158,9 @@ unsafe fn open_vaapi_encoder(
            lp,
        ) {
            Ok(enc) => {
-                LP_MODE[idx].store(if lp { 2 } else { 1 }, Ordering::Relaxed);
+                if let Ok(mut m) = LP_MODE.get_or_init(|| Mutex::new(HashMap::new())).lock() {
+                    m.insert(key.clone(), if lp { 2 } else { 1 });
+                }
                if lp {
                    tracing::info!(
                        encoder = codec.vaapi_name(),
@@ -253,20 +273,20 @@ pub fn probe_can_encode(codec: Codec) -> bool {
    if ffmpeg::init().is_err() {
        return false;
    }
-    // SAFETY: `ffmpeg::init()` returned Ok above, so libav is initialized. `av_log_get_level`/
-    // `av_log_set_level` only read/write libav's global integer log level (no pointer args) and are
-    // always sound to call post-init. `VaapiHw::new` (an `unsafe fn`) builds a VAAPI device + NV12
-    // frames pool from the literal NV12/640x480/pool=2 args and hands back a RAII handle that unrefs
-    // both `AVBufferRef`s on drop. `open_vaapi_encoder` (an `unsafe fn`) borrows `hw.device_ref`/
-    // `hw.frames_ref` — the two non-null refs `VaapiHw::new` just created — and `av_buffer_ref`s them
-    // into the encoder; `hw` is a live local for the whole match arm, so the borrows outlive the
-    // synchronous call, and both `hw` and the probe encoder are dropped (RAII) when the arm ends.
+    // A missing VA device (non-VAAPI host, GPU-less CI) is an expected probe outcome — quiet
+    // ffmpeg's "No VA display found" error for the probe. Held until the function returns, so the
+    // level is restored after the open either way. Shares one lock with the NVENC probes, which
+    // race this one on the same libav global (see [`crate::linux::QuietLibavLog`]).
+    let _quiet = crate::linux::QuietLibavLog::new();
+    // SAFETY: `ffmpeg::init()` returned Ok above, so libav is initialized. `VaapiHw::new` (an
+    // `unsafe fn`) builds a VAAPI device + NV12 frames pool from the literal NV12/640x480/pool=2
+    // args and hands back a RAII handle that unrefs both `AVBufferRef`s on drop.
+    // `open_vaapi_encoder` (an `unsafe fn`) borrows `hw.device_ref`/`hw.frames_ref` — the two
+    // non-null refs `VaapiHw::new` just created — and `av_buffer_ref`s them into the encoder; `hw`
+    // is a live local for the whole match arm, so the borrows outlive the synchronous call, and
+    // both `hw` and the probe encoder are dropped (RAII) when the arm ends.
    unsafe {
-        // A missing VA device (non-VAAPI host, GPU-less CI) is an expected probe outcome — quiet
-        // ffmpeg's "No VA display found" error for the probe, then restore the level.
-        let prev = ffi::av_log_get_level();
-        ffi::av_log_set_level(ffi::AV_LOG_FATAL);
-        let ok = match VaapiHw::new(ffi::AVPixelFormat::AV_PIX_FMT_NV12, 640, 480, 2) {
+        match VaapiHw::new(ffi::AVPixelFormat::AV_PIX_FMT_NV12, 640, 480, 2) {
            Ok(hw) => open_vaapi_encoder(
                codec,
                640,
@@ -279,9 +299,7 @@ pub fn probe_can_encode(codec: Codec) -> bool {
            )
            .is_ok(),
            Err(_) => false,
-        };
-        ffi::av_log_set_level(prev);
-        ok
+        }
    }
 }

@@ -297,18 +315,17 @@ pub fn probe_can_encode_10bit(codec: Codec) -> bool {
    if ffmpeg::init().is_err() {
        return false;
    }
-    // SAFETY: `ffmpeg::init()` returned Ok above, so libav is initialized. `av_log_{get,set}_level`
-    // only read/write libav's global integer log level (no pointer args). `VaapiHw::new` (an
+    // A missing VA device / no Main10 entrypoint is an expected probe outcome — quiet ffmpeg's
+    // error for the probe. Held until the function returns, so the level is restored after the open
+    // either way, and shared with the other probes (see [`crate::linux::QuietLibavLog`]).
+    let _quiet = crate::linux::QuietLibavLog::new();
+    // SAFETY: `ffmpeg::init()` returned Ok above, so libav is initialized. `VaapiHw::new` (an
    // `unsafe fn`) builds a VAAPI device + P010 frames pool from the literal args and hands back a
    // RAII handle; `open_vaapi_encoder` (an `unsafe fn`) borrows `hw.device_ref`/`hw.frames_ref` —
    // the two non-null refs `VaapiHw::new` just created, live locals for the whole match arm — and
    // `av_buffer_ref`s them into the probe encoder. Both `hw` and the encoder drop (RAII) at arm end.
    unsafe {
-        // A missing VA device / no Main10 entrypoint is an expected probe outcome — quiet ffmpeg's
-        // error for the probe, then restore the level.
-        let prev = ffi::av_log_get_level();
-        ffi::av_log_set_level(ffi::AV_LOG_FATAL);
-        let ok = match VaapiHw::new(ffi::AVPixelFormat::AV_PIX_FMT_P010LE, 640, 480, 2) {
+        match VaapiHw::new(ffi::AVPixelFormat::AV_PIX_FMT_P010LE, 640, 480, 2) {
            Ok(hw) => open_vaapi_encoder(
                codec,
                640,
@@ -321,9 +338,7 @@ pub fn probe_can_encode_10bit(codec: Codec) -> bool {
            )
            .is_ok(),
            Err(_) => false,
-        };
-        ffi::av_log_set_level(prev);
-        ok
+        }
    }
 }

@@ -934,7 +949,8 @@ impl DmabufInner {
        //    `Box` puts it on the heap with a unique owner.
        //  * `dmabuf.fd.as_raw_fd()` is the fd of the caller's `&DmabufFrame`, which owns it for the
        //    whole synchronous `submit`; we describe one object/layer/plane from its
-        //    fourcc/modifier/offset/stride and pass `object.size = 0` (ffmpeg queries the real size).
+        //    fourcc/modifier/offset/stride and its `lseek`-queried size. `libc::lseek` on that live
+        //    fd only reads the description's size and returns it (or -1); it touches no Rust memory.
        //  * `av_frame_alloc` → `drm` (null-checked); we set its scalar fields and
        //    `hw_frames_ctx = av_buffer_ref(self.drm_frames)` (new ref of the live owned ctx).
        //  * `data[0] = Box::into_raw(desc)` transfers the box into the frame; `buf[0] =
@@ -950,7 +966,17 @@ impl DmabufInner {
            let mut desc: Box<ffi::AVDRMFrameDescriptor> = Box::new(std::mem::zeroed());
            desc.nb_objects = 1;
            desc.objects[0].fd = dmabuf.fd.as_raw_fd();
-            desc.objects[0].size = 0;
+            // The object's REAL size, not 0. libav does not query it for us — both of its import
+            // paths hand this value straight to libva, as `prime_desc.objects[i].size` on the
+            // PRIME_2 path and `buffer_desc.data_size` on the legacy fallback — so a 0 told every
+            // VA driver the backing object was empty and left it to work the real size out itself.
+            // The drivers this has run on (radeonsi, modern Intel iHD) do; a Gen9 Intel host
+            // answered `vaCreateSurfaces` with VA_STATUS_ERROR_ALLOCATION_FAILED on every single
+            // frame. `lseek(SEEK_END)` is the standard dma-buf size query — the same one the
+            // Vulkan bridge already uses on these fds (`pf_zerocopy::imp::vulkan`). If a kernel
+            // refuses it, keep the old 0 rather than drop a frame we could still have encoded.
+            let obj_size = libc::lseek(dmabuf.fd.as_raw_fd(), 0, libc::SEEK_END);
+            desc.objects[0].size = if obj_size > 0 { obj_size as _ } else { 0 };
            desc.objects[0].format_modifier = dmabuf.modifier;
            desc.nb_layers = 1;
            desc.layers[0].format = self.fourcc;
@@ -999,8 +1025,18 @@ impl DmabufInner {
                ffi::AV_BUFFERSRC_FLAG_KEEP_REF as c_int,
            );
            ffi::av_frame_free(&mut drm);
+            // These two stages ARE the import: the push hands libav our DRM-PRIME descriptor, and
+            // the pull is where `hwmap` actually maps it into a VA surface (and `scale_vaapi` runs
+            // the CSC). A failure here means this driver would not take this compositor's dmabuf —
+            // which no encoder rebuild can fix — so tell the process-wide latch, and capture
+            // negotiates CPU frames from the next session on. `avcodec_send_frame` below is
+            // deliberately NOT counted: that one is the encoder stalling, which the in-place
+            // rebuild above us exists to recover, and disabling zero-copy over it would be a
+            // permanent penalty for a transient fault.
            if r < 0 {
-                bail!("av_buffersrc_add_frame failed ({r})");
+                let e = format!("av_buffersrc_add_frame failed ({r})");
+                pf_zerocopy::note_raw_dmabuf_import_failure(&e);
+                bail!("{e}");
            }
            t_push = t0.elapsed();
            let mut nv12 = ffi::av_frame_alloc();
@@ -1010,8 +1046,11 @@ impl DmabufInner {
            let r = ffi::av_buffersink_get_frame(self.sink, nv12);
            if r < 0 {
                ffi::av_frame_free(&mut nv12);
-                bail!("av_buffersink_get_frame failed ({r})");
+                let e = format!("av_buffersink_get_frame failed ({r})");
+                pf_zerocopy::note_raw_dmabuf_import_failure(&e);
+                bail!("{e}");
            }
+            pf_zerocopy::note_raw_dmabuf_import_ok();
            t_pull = t0.elapsed() - t_push;
            (*nv12).pts = pts;
            (*nv12).pict_type = if idr {
@@ -47,6 +47,15 @@ pub const PRIMARY_REF_NONE: u8 = 7;
 /// `VK_VIDEO_ENCODE_AV1_SUPERBLOCK_SIZE_128_BIT_KHR` (bit 1 of the superblock-size flags).
 pub const SUPERBLOCK_SIZE_128: u32 = 0x2;

+// `VkVideoEncodeAV1CapabilityFlagBitsKHR` — the two that decide whether the encode source may be a
+// different size from the declared frame. Both absent on RADV PHOENIX.
+/// Without this, the source's `codedExtent` MUST equal the sequence header's
+/// `max_frame_{width,height}_minus_1 + 1` (`VUID-vkCmdEncodeVideoKHR-flags-10324`).
+pub const CAPABILITY_FRAME_SIZE_OVERRIDE: u32 = 0x0000_0008;
+/// Without this, EVERY reference slot's `codedExtent` MUST equal the source's
+/// (`VUID-vkCmdEncodeVideoKHR-flags-10325`).
+pub const CAPABILITY_MOTION_VECTOR_SCALING: u32 = 0x0000_0010;
+
 // `VkVideoEncodeAV1PredictionModeKHR`
 pub const PREDICTION_MODE_INTRA_ONLY: i32 = 0;
 pub const PREDICTION_MODE_SINGLE_REFERENCE: i32 = 1;
@@ -498,3 +507,366 @@ pub struct StdVideoEncodeAV1OperatingPointInfo {
 pub fn stype(raw: i32) -> vk::StructureType {
    vk::StructureType::from_raw(raw)
 }
+
+// ---------- ABI layout guard ----------
+//
+// These structs are hand-copied and handed to the driver through raw `p_next` chains, so nothing in
+// the type system relates them to the C definitions any more: an edit that inserts, drops, widens or
+// re-pads a field is not a compile error, it is the driver reading our bytes at the wrong offsets.
+// The assertions below are the missing compile error. They are `const` rather than `#[cfg(test)]`
+// (the shape `amf.rs` uses) so they hold in every build, including the shipped one, and on any
+// target this module compiles for.
+//
+// What they catch: a changed field width, an inserted or removed field, a changed array length, a
+// padding assumption that only holds on one target. What they CANNOT catch: swapping two fields of
+// the same type — offsets are unchanged. That case is only caught by reading the registry, so the
+// field order here was diffed against `vulkan_core.h` and `vk_video/vulkan_video_codec_av1std_encode.h`
+// (Vulkan-Headers `main`, 2026-07-25) when these assertions were written, along with every `ST_*`,
+// flag-bit and enum value above; the bitfield member order is pinned by the test module below.
+//
+// Deliberately duplicated in `vk_valve_rgb.rs` rather than shared: both modules exist to be deleted
+// wholesale once `ash` ships these bindings, and a shared helper would make deleting one break the
+// other.
+macro_rules! assert_abi_layout {
+    ($t:ty { size: $size:expr, align: $align:expr $(, $field:ident @ $off:expr)* $(,)? }) => {
+        const _: () = {
+            assert!(
+                ::core::mem::size_of::<$t>() == $size,
+                concat!(stringify!($t), ": size does not match the C ABI")
+            );
+            assert!(
+                ::core::mem::align_of::<$t>() == $align,
+                concat!(stringify!($t), ": alignment does not match the C ABI")
+            );
+            $(assert!(
+                ::core::mem::offset_of!($t, $field) == $off,
+                concat!(stringify!($t), ".", stringify!($field), ": offset does not match the C ABI")
+            );)*
+        };
+    };
+}
+
+// Std encode structs. The three `*Flags` types are a single C `uint32_t` of bitfields, so only their
+// size and alignment are layout-checkable here; their member order is covered by `abi_tests`.
+assert_abi_layout!(StdVideoEncodeAV1PictureInfoFlags { size: 4, align: 4 });
+
+assert_abi_layout!(StdVideoEncodeAV1PictureInfo {
+    size: 152, align: 8,
+    flags @ 0,
+    frame_type @ 4,
+    frame_presentation_time @ 8,
+    current_frame_id @ 12,
+    order_hint @ 16,
+    primary_ref_frame @ 17,
+    refresh_frame_flags @ 18,
+    coded_denom @ 19,
+    render_width_minus_1 @ 20,
+    render_height_minus_1 @ 22,
+    interpolation_filter @ 24,
+    TxMode @ 28,
+    delta_q_res @ 32,
+    delta_lf_res @ 33,
+    ref_order_hint @ 34,
+    ref_frame_idx @ 42,
+    reserved1 @ 49,
+    delta_frame_id_minus_1 @ 52,
+    pTileInfo @ 80,
+    pQuantization @ 88,
+    pSegmentation @ 96,
+    pLoopFilter @ 104,
+    pCDEF @ 112,
+    pLoopRestoration @ 120,
+    pGlobalMotion @ 128,
+    pExtensionHeader @ 136,
+    pBufferRemovalTimes @ 144,
+});
+
+assert_abi_layout!(StdVideoEncodeAV1ReferenceInfoFlags { size: 4, align: 4 });
+
+assert_abi_layout!(StdVideoEncodeAV1ReferenceInfo {
+    size: 24, align: 8,
+    flags @ 0,
+    RefFrameId @ 4,
+    frame_type @ 8,
+    OrderHint @ 12,
+    reserved1 @ 13,
+    pExtensionHeader @ 16,
+});
+
+assert_abi_layout!(StdVideoEncodeAV1ExtensionHeader {
+    size: 2, align: 1,
+    temporal_id @ 0,
+    spatial_id @ 1,
+});
+
+assert_abi_layout!(StdVideoEncodeAV1OperatingPointInfoFlags { size: 4, align: 4 });
+
+assert_abi_layout!(StdVideoEncodeAV1OperatingPointInfo {
+    size: 20, align: 4,
+    flags @ 0,
+    operating_point_idc @ 4,
+    seq_level_idx @ 6,
+    seq_tier @ 7,
+    decoder_buffer_delay @ 8,
+    encoder_buffer_delay @ 12,
+    initial_display_delay_minus_1 @ 16,
+});
+
+// KHR extension structs.
+assert_abi_layout!(VideoEncodeAV1ProfileInfoKHR {
+    size: 24, align: 8,
+    s_type @ 0,
+    p_next @ 8,
+    std_profile @ 16,
+});
+
+assert_abi_layout!(PhysicalDeviceVideoEncodeAV1FeaturesKHR {
+    size: 24, align: 8,
+    s_type @ 0,
+    p_next @ 8,
+    video_encode_av1 @ 16,
+});
+
+assert_abi_layout!(VideoEncodeAV1CapabilitiesKHR {
+    size: 128, align: 8,
+    s_type @ 0,
+    p_next @ 8,
+    flags @ 16,
+    max_level @ 20,
+    coded_picture_alignment @ 24,
+    max_tiles @ 32,
+    min_tile_size @ 40,
+    max_tile_size @ 48,
+    superblock_sizes @ 56,
+    max_single_reference_count @ 60,
+    single_reference_name_mask @ 64,
+    max_unidirectional_compound_reference_count @ 68,
+    max_unidirectional_compound_group1_reference_count @ 72,
+    unidirectional_compound_reference_name_mask @ 76,
+    max_bidirectional_compound_reference_count @ 80,
+    max_bidirectional_compound_group1_reference_count @ 84,
+    max_bidirectional_compound_group2_reference_count @ 88,
+    bidirectional_compound_reference_name_mask @ 92,
+    max_temporal_layer_count @ 96,
+    max_spatial_layer_count @ 100,
+    max_operating_points @ 104,
+    min_q_index @ 108,
+    max_q_index @ 112,
+    prefers_gop_remaining_frames @ 116,
+    requires_gop_remaining_frames @ 120,
+    std_syntax_flags @ 124,
+});
+
+assert_abi_layout!(VideoEncodeAV1SessionParametersCreateInfoKHR {
+    size: 48, align: 8,
+    s_type @ 0,
+    p_next @ 8,
+    p_std_sequence_header @ 16,
+    p_std_decoder_model_info @ 24,
+    std_operating_point_count @ 32,
+    p_std_operating_points @ 40,
+});
+
+assert_abi_layout!(VideoEncodeAV1PictureInfoKHR {
+    size: 80, align: 8,
+    s_type @ 0,
+    p_next @ 8,
+    prediction_mode @ 16,
+    rate_control_group @ 20,
+    constant_q_index @ 24,
+    p_std_picture_info @ 32,
+    reference_name_slot_indices @ 40,
+    primary_reference_cdf_only @ 68,
+    generate_obu_extension_header @ 72,
+});
+
+assert_abi_layout!(VideoEncodeAV1DpbSlotInfoKHR {
+    size: 24, align: 8,
+    s_type @ 0,
+    p_next @ 8,
+    p_std_reference_info @ 16,
+});
+
+assert_abi_layout!(VideoEncodeAV1RateControlInfoKHR {
+    size: 40, align: 8,
+    s_type @ 0,
+    p_next @ 8,
+    flags @ 16,
+    gop_frame_count @ 20,
+    key_frame_period @ 24,
+    consecutive_bipredictive_frame_count @ 28,
+    temporal_layer_count @ 32,
+});
+
+assert_abi_layout!(VideoEncodeAV1QIndexKHR {
+    size: 12, align: 4,
+    intra_q_index @ 0,
+    predictive_q_index @ 4,
+    bipredictive_q_index @ 8,
+});
+
+assert_abi_layout!(VideoEncodeAV1FrameSizeKHR {
+    size: 12, align: 4,
+    intra_frame_size @ 0,
+    predictive_frame_size @ 4,
+    bipredictive_frame_size @ 8,
+});
+
+assert_abi_layout!(VideoEncodeAV1RateControlLayerInfoKHR {
+    size: 64, align: 8,
+    s_type @ 0,
+    p_next @ 8,
+    use_min_q_index @ 16,
+    min_q_index @ 20,
+    use_max_q_index @ 32,
+    max_q_index @ 36,
+    use_max_frame_size @ 48,
+    max_frame_size @ 52,
+});
+
+assert_abi_layout!(VideoEncodeAV1GopRemainingFrameInfoKHR {
+    size: 32, align: 8,
+    s_type @ 0,
+    p_next @ 8,
+    use_gop_remaining_frames @ 16,
+    gop_remaining_intra @ 20,
+    gop_remaining_predictive @ 24,
+    gop_remaining_bipredictive @ 28,
+});
+
+assert_abi_layout!(VideoEncodeAV1SessionCreateInfoKHR {
+    size: 24, align: 8,
+    s_type @ 0,
+    p_next @ 8,
+    use_max_level @ 16,
+    max_level @ 20,
+});
+
+#[cfg(test)]
+mod abi_tests {
+    use super::*;
+
+    /// Assert that `set` lights exactly one bit of the flags word, at `bit`.
+    fn sets_only_bit(
+        storage: __BindgenBitfieldUnit<[u8; 4]>,
+        name: &str,
+        bit: usize,
+        expected_width: usize,
+    ) {
+        for probe in 0..32 {
+            let want = probe >= bit && probe < bit + expected_width;
+            assert_eq!(
+                storage.get_bit(probe),
+                want,
+                "`{name}` should occupy bit(s) {bit}..{}, but bit {probe} disagrees",
+                bit + expected_width
+            );
+        }
+    }
+
+    /// A C bitfield allocates its members from bit 0 upward in declaration order, so the setter for
+    /// the Nth member of `StdVideoEncodeAV1PictureInfoFlags` must write bit N. The array below is
+    /// the member list of `vulkan_video_codec_av1std_encode.h` **in declaration order** — so this
+    /// pins the hand-copied bit indices to the header rather than merely to themselves. A silent
+    /// renumbering here would make the driver read, say, `use_superres` where we meant
+    /// `render_and_frame_size_different`.
+    #[test]
+    fn picture_info_flag_setters_follow_the_c_declaration_order() {
+        #[allow(clippy::type_complexity)]
+        let members: [(&str, fn(&mut StdVideoEncodeAV1PictureInfoFlags)); 29] = [
+            ("error_resilient_mode", |f| f.set_error_resilient_mode(1)),
+            ("disable_cdf_update", |f| f.set_disable_cdf_update(1)),
+            ("use_superres", |f| f.set_use_superres(1)),
+            ("render_and_frame_size_different", |f| {
+                f.set_render_and_frame_size_different(1)
+            }),
+            ("allow_screen_content_tools", |f| {
+                f.set_allow_screen_content_tools(1)
+            }),
+            ("is_filter_switchable", |f| f.set_is_filter_switchable(1)),
+            ("force_integer_mv", |f| f.set_force_integer_mv(1)),
+            ("frame_size_override_flag", |f| {
+                f.set_frame_size_override_flag(1)
+            }),
+            ("buffer_removal_time_present_flag", |f| {
+                f.set_buffer_removal_time_present_flag(1)
+            }),
+            ("allow_intrabc", |f| f.set_allow_intrabc(1)),
+            ("frame_refs_short_signaling", |f| {
+                f.set_frame_refs_short_signaling(1)
+            }),
+            ("allow_high_precision_mv", |f| {
+                f.set_allow_high_precision_mv(1)
+            }),
+            ("is_motion_mode_switchable", |f| {
+                f.set_is_motion_mode_switchable(1)
+            }),
+            ("use_ref_frame_mvs", |f| f.set_use_ref_frame_mvs(1)),
+            ("disable_frame_end_update_cdf", |f| {
+                f.set_disable_frame_end_update_cdf(1)
+            }),
+            ("allow_warped_motion", |f| f.set_allow_warped_motion(1)),
+            ("reduced_tx_set", |f| f.set_reduced_tx_set(1)),
+            ("skip_mode_present", |f| f.set_skip_mode_present(1)),
+            ("delta_q_present", |f| f.set_delta_q_present(1)),
+            ("delta_lf_present", |f| f.set_delta_lf_present(1)),
+            ("delta_lf_multi", |f| f.set_delta_lf_multi(1)),
+            ("segmentation_enabled", |f| f.set_segmentation_enabled(1)),
+            ("segmentation_update_map", |f| {
+                f.set_segmentation_update_map(1)
+            }),
+            ("segmentation_temporal_update", |f| {
+                f.set_segmentation_temporal_update(1)
+            }),
+            ("segmentation_update_data", |f| {
+                f.set_segmentation_update_data(1)
+            }),
+            ("UsesLr", |f| f.set_UsesLr(1)),
+            ("usesChromaLr", |f| f.set_usesChromaLr(1)),
+            ("show_frame", |f| f.set_show_frame(1)),
+            ("showable_frame", |f| f.set_showable_frame(1)),
+        ];
+
+        for (bit, (name, set)) in members.into_iter().enumerate() {
+            let mut flags = StdVideoEncodeAV1PictureInfoFlags {
+                _bitfield_align_1: [],
+                _bitfield_1: __BindgenBitfieldUnit::new([0u8; 4]),
+            };
+            set(&mut flags);
+            sets_only_bit(flags._bitfield_1, name, bit, 1);
+        }
+    }
+
+    /// `reserved : 3` closes out the word — bits 29..32. Checking it is what proves the 29 members
+    /// above are the *whole* list: a dropped member would shift `reserved` down and fail here.
+    #[test]
+    fn picture_info_reserved_occupies_the_top_three_bits() {
+        let mut flags = StdVideoEncodeAV1PictureInfoFlags {
+            _bitfield_align_1: [],
+            _bitfield_1: __BindgenBitfieldUnit::new([0u8; 4]),
+        };
+        flags.set_reserved(0b111);
+        sets_only_bit(flags._bitfield_1, "reserved", 29, 3);
+    }
+
+    /// The same invariant for the two-member reference-info flags word.
+    #[test]
+    fn reference_info_flag_setters_follow_the_c_declaration_order() {
+        #[allow(clippy::type_complexity)]
+        let members: [(&str, fn(&mut StdVideoEncodeAV1ReferenceInfoFlags)); 2] = [
+            ("disable_frame_end_update_cdf", |f| {
+                f.set_disable_frame_end_update_cdf(1)
+            }),
+            ("segmentation_enabled", |f| f.set_segmentation_enabled(1)),
+        ];
+
+        for (bit, (name, set)) in members.into_iter().enumerate() {
+            let mut flags = StdVideoEncodeAV1ReferenceInfoFlags {
+                _bitfield_align_1: [],
+                _bitfield_1: __BindgenBitfieldUnit::new([0u8; 4]),
+            };
+            set(&mut flags);
+            sets_only_bit(flags._bitfield_1, name, bit, 1);
+        }
+    }
+}
@@ -0,0 +1,836 @@
+//! Session/frame **construction** for the Vulkan Video encoder — the unsafe builders
+//! (`make_frame*`, `make_video_image`, `probe_rgb_direct`) and the parameter-set bitstream
+//! writers (`build_parameters_h265`/`_av1`, the AV1 sequence-header OBU). Split from
+//! `vulkan_video.rs` (WP7.5) the way `amf_sys.rs` was split from `amf.rs`: a `#[path]` child
+//! module, so this file sees the parent's private items (`Frame` and friends) with zero
+//! visibility churn, and ~800 lines of construction `unsafe` get their own review surface.
+//! Steady-state encode logic stays in the parent.
+
+// The parent's whole item namespace (Frame, the consts, sibling helpers) — the point of the
+// child-module shape. External imports are this file's own; `vk_util` is a crate-root sibling,
+// so the path is `crate::`, not the parent-relative `super::` the parent uses.
+use super::*;
+use crate::vk_util::{find_mem, make_plain_image, make_view};
+use anyhow::{bail, Result};
+use ash::vk;
+use std::ffi::c_void;
+
+pub(super) fn align_up(v: u64, a: u64) -> u64 {
+    v.div_ceil(a) * a
+}
+
+/// Probe for the RGB-direct encode source (design/vulkan-rgb-direct-encode.md): can this device
+/// take the captured RGB dmabuf directly, with the VCN EFC front-end doing the 709-narrow CSC,
+/// via `VK_VALVE_video_encode_rgb_conversion` (RADV since Mesa 26.0, gated on EFC hardware)?
+/// `Ok((x_offset, y_offset))` carries the chroma-siting bits a session must be created with
+/// (the preferred available bit per axis); `Err` is the first missing requirement, logged as
+/// the open-time verdict.
+pub(super) unsafe fn probe_rgb_direct(
+    instance: &ash::Instance,
+    vq_inst: &ash::khr::video_queue::Instance,
+    pd: vk::PhysicalDevice,
+    codec_op: vk::VideoCodecOperationFlagsKHR,
+    av1: bool,
+) -> Result<(u32, u32), &'static str> {
+    use crate::vk_av1_encode as av1b;
+    use crate::vk_valve_rgb as vrgb;
+    // 1. The device extension must exist (Mesa >= 26.0 AND the VCN has an EFC block).
+    let Ok(exts) = instance.enumerate_device_extension_properties(pd) else {
+        return Err("probe-failed(ext-enum)");
+    };
+    if !exts
+        .iter()
+        .any(|e| std::ffi::CStr::from_ptr(e.extension_name.as_ptr()) == vrgb::EXTENSION_NAME)
+    {
+        return Err("no-ext(mesa<26.0-or-no-efc)");
+    }
+    // 2. Feature bit.
+    let mut feat = vrgb::PhysicalDeviceVideoEncodeRgbConversionFeaturesVALVE {
+        s_type: vrgb::stype(vrgb::ST_PHYSICAL_DEVICE_FEATURES),
+        p_next: std::ptr::null_mut(),
+        video_encode_rgb_conversion: vk::FALSE,
+    };
+    let mut f2 = vk::PhysicalDeviceFeatures2 {
+        p_next: &mut feat as *mut _ as *mut c_void,
+        ..Default::default()
+    };
+    instance.get_physical_device_features2(pd, &mut f2);
+    if feat.video_encode_rgb_conversion == vk::FALSE {
+        return Err("no-feature");
+    }
+    // 3. Capabilities under the rgb-chained profile — the conversion must cover the compute
+    //    CSC's colour math (rgb2yuv.comp: BT.709, narrow range; chroma siting is looser, see
+    //    below). The profile chain is the same one every rgb-direct consumer presents.
+    let mut ps = RgbProfileStack::new(codec_op);
+    let profile = *ps.wire(av1);
+    let mut rgb_caps = vrgb::VideoEncodeRgbConversionCapabilitiesVALVE {
+        s_type: vrgb::stype(vrgb::ST_CAPABILITIES),
+        p_next: std::ptr::null_mut(),
+        rgb_models: 0,
+        rgb_ranges: 0,
+        x_chroma_offsets: 0,
+        y_chroma_offsets: 0,
+    };
+    let mut h265_caps = vk::VideoEncodeH265CapabilitiesKHR::default();
+    let mut av1_caps: av1b::VideoEncodeAV1CapabilitiesKHR = std::mem::zeroed();
+    av1_caps.s_type = av1b::stype(av1b::ST_CAPABILITIES);
+    let mut enc_caps = vk::VideoEncodeCapabilitiesKHR::default();
+    let mut caps = vk::VideoCapabilitiesKHR::default();
+    if av1 {
+        av1_caps.p_next = &mut rgb_caps as *mut _ as *mut c_void;
+        enc_caps.p_next = &mut av1_caps as *mut _ as *mut c_void;
+    } else {
+        h265_caps.p_next = &mut rgb_caps as *mut _ as *mut c_void;
+        enc_caps.p_next = &mut h265_caps as *mut _ as *mut c_void;
+    }
+    caps.p_next = &mut enc_caps as *mut _ as *mut c_void;
+    let r = (vq_inst.fp().get_physical_device_video_capabilities_khr)(pd, &profile, &mut caps);
+    if r != vk::Result::SUCCESS {
+        return Err("no-rgb-profile(caps)");
+    }
+    // Colour model + range must match the shader exactly (709 narrow). Chroma siting is looser
+    // BY ON-GLASS FINDING (RADV 26.0.4 / 780M): the VCN EFC advertises x=COSITED_EVEN only —
+    // the canonical H.26x left-cosited siting — while our 2x2-average shader is midpoint. The
+    // difference is a half-pel chroma-x phase, imperceptible (and EFC's is arguably the more
+    // correct one since nothing in our bitstream signals siting). Accept either bit per axis
+    // and choose the closest to the shader's math: midpoint if offered, else cosited-even.
+    let pick = |offered: u32| -> Option<u32> {
+        if offered & vrgb::CHROMA_OFFSET_MIDPOINT != 0 {
+            Some(vrgb::CHROMA_OFFSET_MIDPOINT)
+        } else if offered & vrgb::CHROMA_OFFSET_COSITED_EVEN != 0 {
+            Some(vrgb::CHROMA_OFFSET_COSITED_EVEN)
+        } else {
+            None
+        }
+    };
+    if rgb_caps.rgb_models & vrgb::MODEL_YCBCR_709 == 0
+        || rgb_caps.rgb_ranges & vrgb::RANGE_NARROW == 0
+    {
+        return Err("no-709-narrow");
+    }
+    let (Some(x_offset), Some(y_offset)) = (
+        pick(rgb_caps.x_chroma_offsets),
+        pick(rgb_caps.y_chroma_offsets),
+    ) else {
+        return Err("no-chroma-siting");
+    };
+    // 4. The encode-src format set under this profile must offer BGRA with DRM-modifier tiling —
+    //    the capture hands LINEAR BGRx dmabufs (fourcc XR24), which import as B8G8R8A8_UNORM.
+    let profile_arr = [profile];
+    let plist = vk::VideoProfileListInfoKHR::default().profiles(&profile_arr);
+    let mut fmt_info = vk::PhysicalDeviceVideoFormatInfoKHR::default()
+        .image_usage(vk::ImageUsageFlags::VIDEO_ENCODE_SRC_KHR);
+    fmt_info.p_next = &plist as *const _ as *const c_void;
+    let get_fmt = vq_inst.fp().get_physical_device_video_format_properties_khr;
+    let mut count = 0u32;
+    let r = get_fmt(pd, &fmt_info, &mut count, std::ptr::null_mut());
+    if r != vk::Result::SUCCESS || count == 0 {
+        return Err("no-rgb-format");
+    }
+    let mut props = vec![vk::VideoFormatPropertiesKHR::default(); count as usize];
+    let r = get_fmt(pd, &fmt_info, &mut count, props.as_mut_ptr());
+    if r != vk::Result::SUCCESS && r != vk::Result::INCOMPLETE {
+        return Err("no-rgb-format");
+    }
+    if !props[..count as usize].iter().any(|p| {
+        p.format == vk::Format::B8G8R8A8_UNORM
+            && p.image_tiling == vk::ImageTiling::DRM_FORMAT_MODIFIER_EXT
+    }) {
+        return Err("no-bgra-modifier-tiling");
+    }
+    Ok((x_offset, y_offset))
+}
+
+pub(super) unsafe fn make_video_image(
+    device: &ash::Device,
+    mp: &vk::PhysicalDeviceMemoryProperties,
+    fmt: vk::Format,
+    w: u32,
+    h: u32,
+    layers: u32,
+    usage: vk::ImageUsageFlags,
+    profile_list: &mut vk::VideoProfileListInfoKHR,
+    concurrent: &[u32],
+) -> Result<(vk::Image, vk::DeviceMemory)> {
+    let mut ci = vk::ImageCreateInfo::default()
+        .image_type(vk::ImageType::TYPE_2D)
+        .format(fmt)
+        .extent(vk::Extent3D {
+            width: w,
+            height: h,
+            depth: 1,
+        })
+        .mip_levels(1)
+        .array_layers(layers)
+        .samples(vk::SampleCountFlags::TYPE_1)
+        .tiling(vk::ImageTiling::OPTIMAL)
+        .usage(usage)
+        .initial_layout(vk::ImageLayout::UNDEFINED)
+        .push_next(profile_list);
+    if concurrent.len() >= 2 {
+        ci = ci
+            .sharing_mode(vk::SharingMode::CONCURRENT)
+            .queue_family_indices(concurrent);
+    } else {
+        ci = ci.sharing_mode(vk::SharingMode::EXCLUSIVE);
+    }
+    let img = device.create_image(&ci, None)?;
+    let req = device.get_image_memory_requirements(img);
+    // Unwind on failure: callers (the open path) only ever see the completed pair.
+    let mem = match device.allocate_memory(
+        &vk::MemoryAllocateInfo::default()
+            .allocation_size(req.size)
+            .memory_type_index(find_mem(
+                mp,
+                req.memory_type_bits,
+                vk::MemoryPropertyFlags::DEVICE_LOCAL,
+            )),
+        None,
+    ) {
+        Ok(m) => m,
+        Err(e) => {
+            device.destroy_image(img, None);
+            return Err(e.into());
+        }
+    };
+    if let Err(e) = device.bind_image_memory(img, mem, 0) {
+        device.destroy_image(img, None);
+        device.free_memory(mem, None);
+        return Err(e.into());
+    }
+    Ok((img, mem))
+}
+
+/// Build one in-flight frame's private resources: NV12 encode-src, Y/UV CSC scratch, its CSC
+/// descriptor set (Y/UV bound now, RGB per use), the bitstream buffer + feedback query, and the
+/// per-frame command buffers + sync. `profile_list`/`profile` are borrowed only during creation.
+///
+/// Builds in place into `f` — a [`Frame::default`] the caller has already parked in its
+/// [`VkTeardown`] guard — so every handle is owned by the unwind the moment it exists and a
+/// mid-build failure leaks nothing.
+pub(super) unsafe fn make_frame(
+    device: &ash::Device,
+    mem_props: &vk::PhysicalDeviceMemoryProperties,
+    w: u32,
+    h: u32,
+    fams: &[u32],
+    profile: &vk::VideoProfileInfoKHR,
+    profile_list: &mut vk::VideoProfileListInfoKHR,
+    csc_dsl: vk::DescriptorSetLayout,
+    csc_pool: vk::DescriptorPool,
+    cmd_pool: vk::CommandPool,
+    compute_pool: vk::CommandPool,
+    bs_size: u64,
+    sampler: vk::Sampler,
+    with_ts: bool,
+    csc: bool,
+    pad_fmt: Option<vk::Format>,
+    f: &mut Frame,
+) -> Result<()> {
+    // "no cursor uploaded yet" sentinel — a real serial may be 0 (see `prep_cursor`).
+    f.cursor_serial = u64::MAX;
+    // Padded-copy staging (unaligned-mode RGB-direct or native NV12): an aligned encode-src in
+    // the session's picture format, filled by a transfer blit each frame — concurrent compute
+    // (copy) + encode (source read). TRANSFER_SRC because the width-padding pass self-copies the
+    // staging image's own last visible column (see `record_pad_blit`).
+    if let Some(fmt) = pad_fmt {
+        (f.pad_img, f.pad_mem) = make_video_image(
+            device,
+            mem_props,
+            fmt,
+            w,
+            h,
+            1,
+            vk::ImageUsageFlags::VIDEO_ENCODE_SRC_KHR
+                | vk::ImageUsageFlags::TRANSFER_DST
+                | vk::ImageUsageFlags::TRANSFER_SRC,
+            profile_list,
+            fams,
+        )?;
+        f.pad_view = make_view(device, f.pad_img, fmt, 0)?;
+    }
+    // RGB-direct sessions never touch the CSC pipeline: no NV12 encode-src, no Y/UV scratch, no
+    // cursor overlay, no descriptor set — the encode source is the imported RGB itself (or the
+    // CPU staging image, built lazily). Their Frame keeps the null handles (teardown-safe).
+    if csc {
+        make_frame_csc(
+            device,
+            mem_props,
+            w,
+            h,
+            fams,
+            profile_list,
+            csc_dsl,
+            csc_pool,
+            sampler,
+            f,
+        )?;
+    }
+    make_frame_common(
+        device,
+        mem_props,
+        profile,
+        profile_list,
+        cmd_pool,
+        compute_pool,
+        bs_size,
+        with_ts,
+        f,
+    )
+}
+
+/// The CSC-only half of [`make_frame`]: NV12 encode-src + Y/UV scratch + cursor + descriptors.
+#[allow(clippy::too_many_arguments)]
+unsafe fn make_frame_csc(
+    device: &ash::Device,
+    mem_props: &vk::PhysicalDeviceMemoryProperties,
+    w: u32,
+    h: u32,
+    fams: &[u32],
+    profile_list: &mut vk::VideoProfileListInfoKHR,
+    csc_dsl: vk::DescriptorSetLayout,
+    csc_pool: vk::DescriptorPool,
+    sampler: vk::Sampler,
+    f: &mut Frame,
+) -> Result<()> {
+    // NV12 encode-src (filled by the CSC copy) — concurrent compute+encode.
+    (f.nv12_src, f.nv12_mem) = make_video_image(
+        device,
+        mem_props,
+        NV12,
+        w,
+        h,
+        1,
+        vk::ImageUsageFlags::VIDEO_ENCODE_SRC_KHR | vk::ImageUsageFlags::TRANSFER_DST,
+        profile_list,
+        fams,
+    )?;
+    f.nv12_view = make_view(device, f.nv12_src, NV12, 0)?;
+    // CSC scratch (Y R8 full-res, UV RG8 half-res).
+    (f.y_img, f.y_mem, f.y_view) = make_plain_image(
+        device,
+        mem_props,
+        vk::Format::R8_UNORM,
+        w,
+        h,
+        vk::ImageUsageFlags::STORAGE | vk::ImageUsageFlags::TRANSFER_SRC,
+    )?;
+    (f.uv_img, f.uv_mem, f.uv_view) = make_plain_image(
+        device,
+        mem_props,
+        vk::Format::R8G8_UNORM,
+        w / 2,
+        h / 2,
+        vk::ImageUsageFlags::STORAGE | vk::ImageUsageFlags::TRANSFER_SRC,
+    )?;
+    // Cursor overlay: fixed CURSOR_MAX² RGBA8 sampled image + host staging (cursor-as-metadata). The
+    // view/descriptor is static (bound at binding 3 below); only the image *content* changes, and
+    // only when the pointer bitmap does — see `prep_cursor`.
+    (f.cursor_img, f.cursor_mem, f.cursor_view) = make_plain_image(
+        device,
+        mem_props,
+        vk::Format::R8G8B8A8_UNORM,
+        CURSOR_MAX,
+        CURSOR_MAX,
+        vk::ImageUsageFlags::SAMPLED | vk::ImageUsageFlags::TRANSFER_DST,
+    )?;
+    f.cursor_stage = device.create_buffer(
+        &vk::BufferCreateInfo::default()
+            .size((CURSOR_MAX * CURSOR_MAX * 4) as u64)
+            .usage(vk::BufferUsageFlags::TRANSFER_SRC),
+        None,
+    )?;
+    let cs_req = device.get_buffer_memory_requirements(f.cursor_stage);
+    f.cursor_stage_mem = device.allocate_memory(
+        &vk::MemoryAllocateInfo::default()
+            .allocation_size(cs_req.size)
+            .memory_type_index(find_mem(
+                mem_props,
+                cs_req.memory_type_bits,
+                vk::MemoryPropertyFlags::HOST_VISIBLE | vk::MemoryPropertyFlags::HOST_COHERENT,
+            )),
+        None,
+    )?;
+    device.bind_buffer_memory(f.cursor_stage, f.cursor_stage_mem, 0)?;
+    // Descriptor set — Y/UV storage bindings fixed; binding 0 (RGB) rewritten per use; binding 3
+    // (cursor) points at the static cursor image (its layout is SHADER_READ_ONLY once prepped).
+    let dsls = [csc_dsl];
+    f.csc_set = device.allocate_descriptor_sets(
+        &vk::DescriptorSetAllocateInfo::default()
+            .descriptor_pool(csc_pool)
+            .set_layouts(&dsls),
+    )?[0];
+    let y_info = [vk::DescriptorImageInfo::default()
+        .image_view(f.y_view)
+        .image_layout(vk::ImageLayout::GENERAL)];
+    let uv_info = [vk::DescriptorImageInfo::default()
+        .image_view(f.uv_view)
+        .image_layout(vk::ImageLayout::GENERAL)];
+    let cur_info = [vk::DescriptorImageInfo::default()
+        .sampler(sampler)
+        .image_view(f.cursor_view)
+        .image_layout(vk::ImageLayout::SHADER_READ_ONLY_OPTIMAL)];
+    device.update_descriptor_sets(
+        &[
+            vk::WriteDescriptorSet::default()
+                .dst_set(f.csc_set)
+                .dst_binding(1)
+                .descriptor_type(vk::DescriptorType::STORAGE_IMAGE)
+                .image_info(&y_info),
+            vk::WriteDescriptorSet::default()
+                .dst_set(f.csc_set)
+                .dst_binding(2)
+                .descriptor_type(vk::DescriptorType::STORAGE_IMAGE)
+                .image_info(&uv_info),
+            vk::WriteDescriptorSet::default()
+                .dst_set(f.csc_set)
+                .dst_binding(3)
+                .descriptor_type(vk::DescriptorType::COMBINED_IMAGE_SAMPLER)
+                .image_info(&cur_info),
+        ],
+        &[],
+    );
+    Ok(())
+}
+
+/// The mode-independent half of [`make_frame`]: bitstream buffer (+ persistent map), feedback
+/// query, optional timestamp pool, command buffers and sync objects.
+#[allow(clippy::too_many_arguments)]
+unsafe fn make_frame_common(
+    device: &ash::Device,
+    mem_props: &vk::PhysicalDeviceMemoryProperties,
+    profile: &vk::VideoProfileInfoKHR,
+    profile_list: &mut vk::VideoProfileListInfoKHR,
+    cmd_pool: vk::CommandPool,
+    compute_pool: vk::CommandPool,
+    bs_size: u64,
+    with_ts: bool,
+    f: &mut Frame,
+) -> Result<()> {
+    // Bitstream buffer + feedback query.
+    f.bs_buf = device.create_buffer(
+        &vk::BufferCreateInfo::default()
+            .size(bs_size)
+            .usage(vk::BufferUsageFlags::VIDEO_ENCODE_DST_KHR)
+            .push_next(profile_list),
+        None,
+    )?;
+    let bs_req = device.get_buffer_memory_requirements(f.bs_buf);
+    f.bs_mem = device.allocate_memory(
+        &vk::MemoryAllocateInfo::default()
+            .allocation_size(bs_req.size)
+            .memory_type_index(find_mem(
+                mem_props,
+                bs_req.memory_type_bits,
+                vk::MemoryPropertyFlags::HOST_VISIBLE | vk::MemoryPropertyFlags::HOST_COHERENT,
+            )),
+        None,
+    )?;
+    device.bind_buffer_memory(f.bs_buf, f.bs_mem, 0)?;
+    // Map once for the slot's lifetime — read_slot copies AUs straight out of this (coherent
+    // memory, no per-frame map/unmap); vkFreeMemory implicitly unmaps at teardown.
+    f.bs_ptr = BsPtr(
+        device.map_memory(f.bs_mem, 0, vk::WHOLE_SIZE, vk::MemoryMapFlags::empty())? as *const u8,
+    );
+    // PUNKTFUNK_PERF: a 2-slot timestamp pool bracketing this slot's compute batch (CSC split).
+    if with_ts {
+        f.ts_pool = device.create_query_pool(
+            &vk::QueryPoolCreateInfo::default()
+                .query_type(vk::QueryType::TIMESTAMP)
+                .query_count(2),
+            None,
+        )?;
+    }
+    let mut fb_ci = vk::QueryPoolVideoEncodeFeedbackCreateInfoKHR::default().encode_feedback_flags(
+        vk::VideoEncodeFeedbackFlagsKHR::BITSTREAM_BUFFER_OFFSET
+            | vk::VideoEncodeFeedbackFlagsKHR::BITSTREAM_BYTES_WRITTEN,
+    );
+    fb_ci.p_next = profile as *const _ as *const c_void;
+    let mut query_ci = vk::QueryPoolCreateInfo::default()
+        .query_type(vk::QueryType::VIDEO_ENCODE_FEEDBACK_KHR)
+        .query_count(1);
+    query_ci.p_next = &fb_ci as *const _ as *const c_void;
+    f.query_pool = device.create_query_pool(&query_ci, None)?;
+    // Command buffers + per-frame sync.
+    f.cmd = device.allocate_command_buffers(
+        &vk::CommandBufferAllocateInfo::default()
+            .command_pool(cmd_pool)
+            .command_buffer_count(1),
+    )?[0];
+    f.compute_cmd = device.allocate_command_buffers(
+        &vk::CommandBufferAllocateInfo::default()
+            .command_pool(compute_pool)
+            .command_buffer_count(1),
+    )?[0];
+    f.csc_sem = device.create_semaphore(&vk::SemaphoreCreateInfo::default(), None)?;
+    f.fence = device.create_fence(&vk::FenceCreateInfo::default(), None)?;
+    Ok(())
+}
+
+/// Author VPS/SPS/PPS (Main, level 4.0, low-latency, conformance-window crop) and return the
+/// session-parameters object + the encoded header bytes (VPS+SPS+PPS NALs) for keyframes.
+pub(super) unsafe fn build_parameters_h265(
+    device: &ash::Device,
+    vq_dev: &ash::khr::video_queue::Device,
+    venc_dev: &ash::khr::video_encode_queue::Device,
+    session: vk::VideoSessionKHR,
+    w: u32,
+    h: u32,
+    rw: u32,
+    rh: u32,
+    quality_level: u32,
+) -> Result<(vk::VideoSessionParametersKHR, Vec<u8>)> {
+    use ash::vk::native as hh;
+    let mut ptl: hh::StdVideoH265ProfileTierLevel = std::mem::zeroed();
+    ptl.flags.set_general_progressive_source_flag(1);
+    ptl.flags.set_general_frame_only_constraint_flag(1);
+    ptl.general_profile_idc = hh::StdVideoH265ProfileIdc_STD_VIDEO_H265_PROFILE_IDC_MAIN;
+    ptl.general_level_idc = hh::StdVideoH265LevelIdc_STD_VIDEO_H265_LEVEL_IDC_6_0;
+
+    let mut dpbm: hh::StdVideoH265DecPicBufMgr = std::mem::zeroed();
+    dpbm.max_dec_pic_buffering_minus1[0] = (DPB_SLOTS - 1) as u8;
+    dpbm.max_num_reorder_pics[0] = 0;
+    dpbm.max_latency_increase_plus1[0] = 0;
+
+    let mut vps: hh::StdVideoH265VideoParameterSet = std::mem::zeroed();
+    vps.flags.set_vps_temporal_id_nesting_flag(1);
+    vps.flags.set_vps_sub_layer_ordering_info_present_flag(1);
+    vps.pDecPicBufMgr = &dpbm;
+    vps.pProfileTierLevel = &ptl;
+
+    let mut sps: hh::StdVideoH265SequenceParameterSet = std::mem::zeroed();
+    sps.flags.set_sps_temporal_id_nesting_flag(1);
+    sps.flags.set_sps_sub_layer_ordering_info_present_flag(1);
+    sps.chroma_format_idc = hh::StdVideoH265ChromaFormatIdc_STD_VIDEO_H265_CHROMA_FORMAT_IDC_420;
+    sps.pic_width_in_luma_samples = w;
+    sps.pic_height_in_luma_samples = h;
+    sps.log2_max_pic_order_cnt_lsb_minus4 = 4;
+    sps.log2_diff_max_min_luma_coding_block_size = 3;
+    sps.log2_diff_max_min_luma_transform_block_size = 3;
+    sps.max_transform_hierarchy_depth_inter = 4;
+    sps.max_transform_hierarchy_depth_intra = 4;
+    sps.pProfileTierLevel = &ptl;
+    sps.pDecPicBufMgr = &dpbm;
+    if w != rw || h != rh {
+        sps.flags.set_conformance_window_flag(1);
+        sps.conf_win_right_offset = (w - rw) / 2; // 4:2:0 SubWidthC = 2
+        sps.conf_win_bottom_offset = (h - rh) / 2; // 4:2:0 SubHeightC = 2
+    }
+
+    let mut pps: hh::StdVideoH265PictureParameterSet = std::mem::zeroed();
+    pps.flags.set_cu_qp_delta_enabled_flag(1);
+    pps.flags.set_pps_loop_filter_across_slices_enabled_flag(1);
+
+    let vps_arr = [vps];
+    let sps_arr = [sps];
+    let pps_arr = [pps];
+    let add = vk::VideoEncodeH265SessionParametersAddInfoKHR::default()
+        .std_vp_ss(&vps_arr)
+        .std_sp_ss(&sps_arr)
+        .std_pp_ss(&pps_arr);
+    let mut h265_ci = vk::VideoEncodeH265SessionParametersCreateInfoKHR::default()
+        .max_std_vps_count(1)
+        .max_std_sps_count(1)
+        .max_std_pps_count(1)
+        .parameters_add_info(&add);
+    // Bake the session's quality level into the parameters object — the spec requires it to match
+    // the level the first frame's ENCODE_QUALITY_LEVEL control installs.
+    let mut q_info = vk::VideoEncodeQualityLevelInfoKHR::default().quality_level(quality_level);
+    let ci = vk::VideoSessionParametersCreateInfoKHR::default()
+        .video_session(session)
+        .push_next(&mut h265_ci)
+        .push_next(&mut q_info);
+    let mut params = vk::VideoSessionParametersKHR::null();
+    let r = (vq_dev.fp().create_video_session_parameters_khr)(
+        device.handle(),
+        &ci,
+        std::ptr::null(),
+        &mut params,
+    );
+    if r != vk::Result::SUCCESS {
+        bail!("create_video_session_parameters: {r:?}");
+    }
+
+    let mut get_h265 = vk::VideoEncodeH265SessionParametersGetInfoKHR::default()
+        .write_std_vps(true)
+        .write_std_sps(true)
+        .write_std_pps(true)
+        .std_vps_id(0)
+        .std_sps_id(0)
+        .std_pps_id(0);
+    let get = vk::VideoEncodeSessionParametersGetInfoKHR::default()
+        .video_session_parameters(params)
+        .push_next(&mut get_h265);
+    let get_fn = venc_dev.fp().get_encoded_video_session_parameters_khr;
+    let mut fb = vk::VideoEncodeSessionParametersFeedbackInfoKHR::default();
+    let mut size: usize = 0;
+    let r = get_fn(
+        device.handle(),
+        &get,
+        &mut fb,
+        &mut size,
+        std::ptr::null_mut(),
+    );
+    if r != vk::Result::SUCCESS {
+        // `params` is live but not yet the caller's guard's to unwind — destroy before bailing.
+        (vq_dev.fp().destroy_video_session_parameters_khr)(
+            device.handle(),
+            params,
+            std::ptr::null(),
+        );
+        bail!("get header size: {r:?}");
+    }
+    let mut buf = vec![0u8; size];
+    let r = get_fn(
+        device.handle(),
+        &get,
+        &mut fb,
+        &mut size,
+        buf.as_mut_ptr() as *mut c_void,
+    );
+    if r != vk::Result::SUCCESS {
+        (vq_dev.fp().destroy_video_session_parameters_khr)(
+            device.handle(),
+            params,
+            std::ptr::null(),
+        );
+        bail!("get header bytes: {r:?}");
+    }
+    buf.truncate(size);
+    Ok((params, buf))
+}
+
+/// AV1 low-overhead OBU bit-writer (MSB-first), used to hand-pack the sequence-header OBU that
+/// Vulkan AV1 encode (unlike H26x) never emits itself.
+struct Av1BitWriter {
+    buf: Vec<u8>,
+    cur: u8,
+    fill: u8,
+}
+impl Av1BitWriter {
+    fn new() -> Self {
+        Self {
+            buf: Vec::new(),
+            cur: 0,
+            fill: 0,
+        }
+    }
+    fn bit(&mut self, b: u32) {
+        self.cur = (self.cur << 1) | (b as u8 & 1);
+        self.fill += 1;
+        if self.fill == 8 {
+            self.buf.push(self.cur);
+            self.cur = 0;
+            self.fill = 0;
+        }
+    }
+    fn put(&mut self, val: u32, bits: u32) {
+        for i in (0..bits).rev() {
+            self.bit((val >> i) & 1);
+        }
+    }
+    /// Flush, zero-padding the final partial byte (OBU size field delimits the payload).
+    fn finish(mut self) -> Vec<u8> {
+        if self.fill > 0 {
+            self.cur <<= 8 - self.fill;
+            self.buf.push(self.cur);
+        }
+        self.buf
+    }
+}
+
+/// AV1 leb128 (little-endian base-128) encoding of an OBU size.
+fn leb128(mut v: u64) -> Vec<u8> {
+    let mut out = Vec::new();
+    loop {
+        let mut byte = (v & 0x7f) as u8;
+        v >>= 7;
+        if v != 0 {
+            byte |= 0x80;
+        }
+        out.push(byte);
+        if v == 0 {
+            break;
+        }
+    }
+    out
+}
+
+/// Bit-pack a `sequence_header_obu` (AV1 spec §5.5) into a size-delimited OBU. The field values here
+/// MUST mirror the `StdVideoAV1SequenceHeader` handed to the driver in `build_parameters_av1` so the
+/// driver-emitted frame OBUs parse against this header. Single operating point, 8-bit 4:2:0,
+/// order-hint on, CDEF+restoration+filter-intra allowed, everything exotic (compound/warp/superres)
+/// disabled — the profile our single-reference P-frame encoder actually uses.
+fn av1_sequence_header_obu(
+    sb128: bool,
+    fwb: u32,
+    fhb: u32,
+    max_w_m1: u32,
+    max_h_m1: u32,
+    order_hint_bits_minus_1: u32,
+    seq_level_idx: u32,
+) -> Vec<u8> {
+    let mut w = Av1BitWriter::new();
+    w.put(0, 3); // seq_profile = MAIN
+    w.bit(0); // still_picture
+    w.bit(0); // reduced_still_picture_header
+    w.bit(0); // timing_info_present_flag
+    w.bit(0); // initial_display_delay_present_flag
+    w.put(0, 5); // operating_points_cnt_minus_1 = 0
+    w.put(0, 12); // operating_point_idc[0]
+    w.put(seq_level_idx, 5); // seq_level_idx[0]
+    if seq_level_idx > 7 {
+        w.bit(0); // seq_tier[0] = 0
+    }
+    w.put(fwb, 4); // frame_width_bits_minus_1
+    w.put(fhb, 4); // frame_height_bits_minus_1
+    w.put(max_w_m1, fwb + 1); // max_frame_width_minus_1
+    w.put(max_h_m1, fhb + 1); // max_frame_height_minus_1
+    w.bit(0); // frame_id_numbers_present_flag
+    w.bit(sb128 as u32); // use_128x128_superblock
+    w.bit(0); // enable_filter_intra
+    w.bit(0); // enable_intra_edge_filter
+    w.bit(0); // enable_interintra_compound
+    w.bit(0); // enable_masked_compound
+    w.bit(0); // enable_warped_motion
+    w.bit(0); // enable_dual_filter
+    w.bit(1); // enable_order_hint
+    w.bit(0); // enable_jnt_comp
+    w.bit(0); // enable_ref_frame_mvs
+    w.bit(1); // seq_choose_screen_content_tools -> seq_force_screen_content_tools = SELECT
+    w.bit(1); // seq_choose_integer_mv -> seq_force_integer_mv = SELECT
+    w.put(order_hint_bits_minus_1, 3); // order_hint_bits_minus_1
+    w.bit(0); // enable_superres
+    w.bit(0); // enable_cdef
+    w.bit(0); // enable_restoration
+              // color_config(): 8-bit 4:2:0, unspecified primaries/transfer/matrix, limited range
+    w.bit(0); // high_bitdepth
+    w.bit(0); // mono_chrome
+    w.bit(0); // color_description_present_flag
+    w.bit(0); // color_range (studio/limited)
+    w.put(0, 2); // chroma_sample_position = CSP_UNKNOWN (subsampling_x==subsampling_y==1 for profile 0)
+    w.bit(0); // separate_uv_delta_q
+    w.bit(0); // film_grain_params_present
+
+    // trailing_bits(): a stop `1` bit then zero-pad to a byte (the size field delimits the OBU, but
+    // the parser still requires the trailing_one_bit — dav1d/cbs reject a plain zero pad).
+    w.bit(1);
+    let payload = w.finish();
+    let mut obu = vec![0x0au8]; // obu_header: type=OBU_SEQUENCE_HEADER(1), has_size_field=1
+    obu.extend_from_slice(&leb128(payload.len() as u64));
+    obu.extend_from_slice(&payload);
+    obu
+}
+
+/// AV1 session parameters + header framing. Vulkan AV1 encode emits only the per-frame OBU, so we
+/// return the app-owned prefixes: a temporal-delimiter OBU that opens every temporal unit
+/// (`frame_prefix`), and TD + the bit-packed sequence-header OBU for keyframes (`header`).
+#[allow(clippy::too_many_arguments)]
+pub(super) unsafe fn build_parameters_av1(
+    device: &ash::Device,
+    vq_dev: &ash::khr::video_queue::Device,
+    session: vk::VideoSessionKHR,
+    w: u32,
+    h: u32,
+    _rw: u32,
+    _rh: u32,
+    max_level: ash::vk::native::StdVideoAV1Level,
+    sb128: bool,
+    quality_level: u32,
+) -> Result<(vk::VideoSessionParametersKHR, Vec<u8>, Vec<u8>)> {
+    use crate::vk_av1_encode as av1;
+    use ash::vk::native as hh;
+
+    let fwb = 31 - w.leading_zeros(); // av_log2(w): enough bits for max_frame_width_minus_1 = w-1
+    let fhb = 31 - h.leading_zeros();
+    let order_hint_bits_minus_1: u32 = 7; // OrderHintBits = 8
+    let seq_level_idx = max_level; // StdVideoAV1Level's numeric value IS the AV1 seq_level_idx
+
+    // ---- Std sequence header (must match the OBU packed below) ----
+    let mut cc_flags: hh::StdVideoAV1ColorConfigFlags = std::mem::zeroed();
+    let _ = &mut cc_flags; // all zero: mono_chrome/color_range/description/separate_uv_delta_q = 0
+    let mut cc: hh::StdVideoAV1ColorConfig = std::mem::zeroed();
+    cc.flags = cc_flags;
+    cc.BitDepth = 8;
+    cc.subsampling_x = 1;
+    cc.subsampling_y = 1;
+    cc.color_primaries = hh::StdVideoAV1ColorPrimaries_STD_VIDEO_AV1_COLOR_PRIMARIES_BT_UNSPECIFIED;
+    cc.transfer_characteristics =
+        hh::StdVideoAV1TransferCharacteristics_STD_VIDEO_AV1_TRANSFER_CHARACTERISTICS_UNSPECIFIED;
+    cc.matrix_coefficients =
+        hh::StdVideoAV1MatrixCoefficients_STD_VIDEO_AV1_MATRIX_COEFFICIENTS_UNSPECIFIED;
+    cc.chroma_sample_position =
+        hh::StdVideoAV1ChromaSamplePosition_STD_VIDEO_AV1_CHROMA_SAMPLE_POSITION_UNKNOWN;
+
+    // Match FFmpeg's Vulkan AV1 encoder (proven on this RADV/VCN path): the ONLY coding tools
+    // enabled are order-hint and (per caps) 128x128 superblocks. CDEF, loop restoration, filter-
+    // intra, warped/compound motion, superres all OFF — enabling them made VCN emit frame-header
+    // sections whose bit layout our sequence header didn't match, desyncing every inter frame.
+    let mut sh_flags: hh::StdVideoAV1SequenceHeaderFlags = std::mem::zeroed();
+    if sb128 {
+        sh_flags.set_use_128x128_superblock(1);
+    }
+    sh_flags.set_enable_order_hint(1);
+    let mut sh: hh::StdVideoAV1SequenceHeader = std::mem::zeroed();
+    sh.flags = sh_flags;
+    sh.seq_profile = hh::StdVideoAV1Profile_STD_VIDEO_AV1_PROFILE_MAIN;
+    sh.frame_width_bits_minus_1 = fwb as u8;
+    sh.frame_height_bits_minus_1 = fhb as u8;
+    sh.max_frame_width_minus_1 = (w - 1) as u16;
+    sh.max_frame_height_minus_1 = (h - 1) as u16;
+    sh.order_hint_bits_minus_1 = order_hint_bits_minus_1 as u8;
+    sh.seq_force_integer_mv = 2; // SELECT
+    sh.seq_force_screen_content_tools = 2; // SELECT
+    sh.pColorConfig = &cc;
+
+    // ---- single operating point conveying the level/tier the driver targets ----
+    let op = av1::StdVideoEncodeAV1OperatingPointInfo {
+        flags: std::mem::zeroed(),
+        operating_point_idc: 0,
+        seq_level_idx: seq_level_idx as u8,
+        seq_tier: 0,
+        decoder_buffer_delay: 0,
+        encoder_buffer_delay: 0,
+        initial_display_delay_minus_1: 0,
+    };
+    let ops = [op];
+    let av1_spci = av1::VideoEncodeAV1SessionParametersCreateInfoKHR {
+        s_type: av1::stype(av1::ST_SESSION_PARAMETERS_CREATE_INFO),
+        p_next: std::ptr::null(),
+        p_std_sequence_header: &sh,
+        p_std_decoder_model_info: std::ptr::null(),
+        std_operating_point_count: 1,
+        p_std_operating_points: ops.as_ptr() as *const c_void,
+    };
+    // Bake the session's quality level into the parameters object (must match the level the first
+    // frame's ENCODE_QUALITY_LEVEL control installs); chained raw ahead of the vendored AV1 struct.
+    let mut q_info = vk::VideoEncodeQualityLevelInfoKHR::default().quality_level(quality_level);
+    q_info.p_next = &av1_spci as *const _ as *const c_void;
+    let mut ci = vk::VideoSessionParametersCreateInfoKHR::default().video_session(session);
+    ci.p_next = &q_info as *const _ as *const c_void;
+    let mut params = vk::VideoSessionParametersKHR::null();
+    let r = (vq_dev.fp().create_video_session_parameters_khr)(
+        device.handle(),
+        &ci,
+        std::ptr::null(),
+        &mut params,
+    );
+    if r != vk::Result::SUCCESS {
+        bail!("create_video_session_parameters (av1): {r:?}");
+    }
+
+    // ---- header framing: TD every temporal unit; TD + seq-header OBU on keyframes ----
+    let td = vec![0x12u8, 0x00]; // temporal_delimiter OBU (type=2, size=0)
+    let seq_obu = av1_sequence_header_obu(
+        sb128,
+        fwb,
+        fhb,
+        w - 1,
+        h - 1,
+        order_hint_bits_minus_1,
+        seq_level_idx,
+    );
+    let mut keyframe_prefix = td.clone();
+    keyframe_prefix.extend_from_slice(&seq_obu);
+    Ok((params, keyframe_prefix, td))
+}
@@ -54,6 +54,59 @@ pub(crate) fn pixel_to_vk(fmt: PixelFormat) -> Option<vk::Format> {
    }
 }

+/// Normalize a CPU RGB payload for Vulkan upload. The packed 24-bpp `Rgb`/`Bgr` the PipeWire
+/// capturer can negotiate are expanded 3→4 into `scratch` (kept by the caller across frames — no
+/// per-frame allocation) with the pad byte = 0xFF; refusing them instead used to kill a session
+/// at its first frame (WP5.4). No packed 24-bpp VkFormat is reliably uploadable/sampleable on
+/// target GPUs, and this path is CPU-sourced by definition, so one cheap expand pass serves it
+/// (the same call NVENC answers with its swscale 3→4 expand, WP1.4).
+///
+/// `bgra_target = false` (the CSC paths): channel order is preserved — the sampler reads through
+/// the matching view format, so any 4-bpp order works and 4-bpp inputs pass through borrowed.
+/// `bgra_target = true` (the RGB-direct encode source): the output byte order is forced to
+/// B,G,R,X, because the video session's `pictureFormat` is `B8G8R8A8_UNORM` and
+/// VUID-vkCmdEncodeVideoKHR-pEncodeInfo-08207 requires the source image to match it — an
+/// R-first source (`Rgbx`/`Rgba`/`Rgb`) is channel-swapped during the same pass. (Caught live on
+/// RADV by `vulkan_smoke_rgb_cpu24`; the mismatch predates the 24-bpp support for `Rgbx` CPU
+/// sources.)
+///
+/// Payloads are tightly packed with no row padding (`FramePayload::Cpu`'s contract), so the
+/// conversion is row-agnostic; a truncated source yields a truncated output, which the upload
+/// paths already bound-check exactly as they did the raw bytes.
+pub(crate) fn normalize_cpu_rgb<'a>(
+    fmt: PixelFormat,
+    bytes: &'a [u8],
+    scratch: &'a mut Vec<u8>,
+    bgra_target: bool,
+) -> (PixelFormat, &'a [u8]) {
+    // Per-pixel source layout: bytes-per-pixel + where R, G, B sit in each pixel.
+    let (bpp, r, g, b) = match fmt {
+        PixelFormat::Rgb => (3usize, 0usize, 1usize, 2usize),
+        PixelFormat::Bgr => (3, 2, 1, 0),
+        PixelFormat::Rgbx | PixelFormat::Rgba => (4, 0, 1, 2),
+        PixelFormat::Bgrx | PixelFormat::Bgra => (4, 2, 1, 0),
+        _ => return (fmt, bytes),
+    };
+    if bpp == 4 && (!bgra_target || b == 0) {
+        return (fmt, bytes); // 4-bpp in an acceptable order: borrow untouched
+    }
+    let px = bytes.len() / bpp;
+    scratch.clear();
+    scratch.resize(px * 4, 0xFF);
+    let (dr, dg, db) = if bgra_target { (2, 1, 0) } else { (r, g, b) };
+    for (dst, src) in scratch.chunks_exact_mut(4).zip(bytes.chunks_exact(bpp)) {
+        dst[dr] = src[r];
+        dst[dg] = src[g];
+        dst[db] = src[b];
+    }
+    let out_fmt = if bgra_target || b == 0 {
+        PixelFormat::Bgrx
+    } else {
+        PixelFormat::Rgbx
+    };
+    (out_fmt, scratch.as_slice())
+}
+
 pub(crate) unsafe fn make_view(
    device: &ash::Device,
    image: vk::Image,
@@ -70,6 +123,21 @@ pub(crate) unsafe fn make_view(
    )?)
 }

+/// Whether a failed dmabuf import should count toward pf-zerocopy's raw-dmabuf degrade latch
+/// (`note_raw_dmabuf_import_failure` — 3 consecutive failures flip capture to CPU delivery for
+/// the process). Deterministic refusals (unsupported fourcc, the driver rejecting the buffer)
+/// must count — they repeat identically forever and the latch is their only recovery. Transient
+/// VRAM pressure must NOT: three tight allocation OOMs would otherwise permanently downgrade a
+/// working host to CPU capture.
+pub(crate) fn import_failure_feeds_latch(e: &anyhow::Error) -> bool {
+    match e.downcast_ref::<vk::Result>() {
+        Some(&r) => {
+            r != vk::Result::ERROR_OUT_OF_DEVICE_MEMORY && r != vk::Result::ERROR_OUT_OF_HOST_MEMORY
+        }
+        None => true,
+    }
+}
+
 /// Import a packed-RGB dmabuf as a SAMPLED VkImage (explicit DRM modifier). Caller destroys all
 /// three returned handles. Extracted verbatim from `vulkan_video.rs`'s import path.
 pub(crate) unsafe fn import_rgb_dmabuf(
@@ -108,9 +176,15 @@ pub(crate) unsafe fn import_rgb_dmabuf_as(
    profile_list: Option<&mut vk::VideoProfileListInfoKHR>,
 ) -> Result<(vk::Image, vk::DeviceMemory, vk::ImageView)> {
    use anyhow::Context;
-    use std::os::fd::IntoRawFd;
+    use std::os::fd::{AsRawFd, IntoRawFd};
    let fmt = fourcc_to_vk(d.fourcc)
        .with_context(|| format!("unsupported dmabuf fourcc {:#x}", d.fourcc))?;
+    // Dup the fd FIRST, and keep it OWNED: ownership transfers to Vulkan only on a SUCCESSFUL
+    // `allocate_memory` (VK_KHR_external_memory_fd — from then on `vkFreeMemory` closes it), so
+    // the release below sits in exactly that arm. Every earlier failure drops the `OwnedFd` for
+    // a single clean close. An explicit `close` after a successful import would be a double
+    // close — and a recycled fd number then clobbers an unrelated descriptor in this process.
+    let dup = d.fd.try_clone().context("dup dmabuf fd")?;
    let planes: Vec<vk::SubresourceLayout> = if fmt == vk::Format::G8_B8R8_2PLANE_420_UNORM {
        let (uv_offset, uv_stride) = d.plane1.map(|(o, s)| (o as u64, s as u64)).unwrap_or((
            d.offset as u64 + d.stride as u64 * ch as u64,
@@ -155,14 +229,16 @@ pub(crate) unsafe fn import_rgb_dmabuf_as(
        ci = ci.push_next(pl);
    }
    let img = device.create_image(&ci, None)?;
-    // dup the fd; Vulkan takes ownership of the dup on a successful import.
-    let dup = d.fd.try_clone().context("dup dmabuf fd")?.into_raw_fd();
+    // Unwind discipline below mirrors `make_plain_image`: every failure destroys what this call
+    // created (and ONLY that — the caller's `DmabufFrame` fd stays theirs).
    let fd_props = {
        let mut p = vk::MemoryFdPropertiesKHR::default();
+        // Borrow-only query (no ownership transfer); an error leaves memory_type_bits = 0 and
+        // the fallback below uses the image requirements alone.
        let _ = (ext_fd.fp().get_memory_fd_properties_khr)(
            device.handle(),
            vk::ExternalMemoryHandleTypeFlags::DMA_BUF_EXT,
-            dup,
+            dup.as_raw_fd(),
            &mut p,
        );
        p.memory_type_bits
@@ -181,27 +257,87 @@ pub(crate) unsafe fn import_rgb_dmabuf_as(
    let mut ded = vk::MemoryDedicatedAllocateInfo::default().image(img);
    let mut import = vk::ImportMemoryFdInfoKHR::default()
        .handle_type(vk::ExternalMemoryHandleTypeFlags::DMA_BUF_EXT)
-        .fd(dup);
-    let mem = device.allocate_memory(
+        .fd(dup.as_raw_fd());
+    let mem = match device.allocate_memory(
        &vk::MemoryAllocateInfo::default()
            .allocation_size(req.size)
            .memory_type_index(ti)
            .push_next(&mut ded)
            .push_next(&mut import),
        None,
-    )?;
-    device.bind_image_memory(img, mem, 0)?;
-    let view = device.create_image_view(
+    ) {
+        Ok(mem) => {
+            // Success transferred fd ownership to the memory object — release, don't close.
+            let _ = dup.into_raw_fd();
+            mem
+        }
+        Err(e) => {
+            device.destroy_image(img, None);
+            return Err(e.into()); // `dup` drops here: the one close of the failed import's fd
+        }
+    };
+    if let Err(e) = device.bind_image_memory(img, mem, 0) {
+        device.destroy_image(img, None);
+        device.free_memory(mem, None); // closes the imported fd
+        return Err(e.into());
+    }
+    let view = match device.create_image_view(
        &vk::ImageViewCreateInfo::default()
            .image(img)
            .view_type(vk::ImageViewType::TYPE_2D)
            .format(fmt)
            .subresource_range(color_range(0)),
        None,
-    )?;
+    ) {
+        Ok(v) => v,
+        Err(e) => {
+            device.destroy_image(img, None);
+            device.free_memory(mem, None);
+            return Err(e.into());
+        }
+    };
    Ok((img, mem, view))
 }

+/// Create + allocate + bind a host-visible/coherent buffer with `make_plain_image`'s unwind
+/// discipline: on any failure everything this call created is destroyed before returning, so
+/// callers can `?` freely. Both `ensure_cpu_rgb` staging twins open-coded this sequence and
+/// leaked the buffer (and then buffer+memory) on the allocate/bind failure arms.
+pub(crate) unsafe fn make_host_buffer(
+    device: &ash::Device,
+    mp: &vk::PhysicalDeviceMemoryProperties,
+    size: u64,
+    usage: vk::BufferUsageFlags,
+) -> Result<(vk::Buffer, vk::DeviceMemory)> {
+    let buf = device.create_buffer(
+        &vk::BufferCreateInfo::default().size(size).usage(usage),
+        None,
+    )?;
+    let req = device.get_buffer_memory_requirements(buf);
+    let mem = match device.allocate_memory(
+        &vk::MemoryAllocateInfo::default()
+            .allocation_size(req.size)
+            .memory_type_index(find_mem(
+                mp,
+                req.memory_type_bits,
+                vk::MemoryPropertyFlags::HOST_VISIBLE | vk::MemoryPropertyFlags::HOST_COHERENT,
+            )),
+        None,
+    ) {
+        Ok(m) => m,
+        Err(e) => {
+            device.destroy_buffer(buf, None);
+            return Err(e.into());
+        }
+    };
+    if let Err(e) = device.bind_buffer_memory(buf, mem, 0) {
+        device.destroy_buffer(buf, None);
+        device.free_memory(mem, None);
+        return Err(e.into());
+    }
+    Ok((buf, mem))
+}
+
 pub(crate) unsafe fn make_plain_image(
    device: &ash::Device,
    mp: &vk::PhysicalDeviceMemoryProperties,
@@ -259,3 +395,80 @@ pub(crate) unsafe fn make_plain_image(
        }
    }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// CSC mode (`bgra_target = false`): the 3→4 expand is a pure byte shuffle — no channel
+    /// reorder, pad byte 0xFF, truncated tail pixels dropped (never overrun) — and 4-bpp inputs
+    /// pass through borrowed untouched.
+    #[test]
+    fn normalize_cpu_rgb_expands_24bpp_and_borrows_4bpp() {
+        let mut scratch = Vec::new();
+        let (f, b) = normalize_cpu_rgb(PixelFormat::Rgb, &[1, 2, 3, 4, 5, 6], &mut scratch, false);
+        assert_eq!(f, PixelFormat::Rgbx);
+        assert_eq!(b, &[1, 2, 3, 0xFF, 4, 5, 6, 0xFF]);
+
+        let mut scratch = Vec::new();
+        let (f, b) = normalize_cpu_rgb(PixelFormat::Bgr, &[9, 8, 7], &mut scratch, false);
+        assert_eq!(f, PixelFormat::Bgrx);
+        assert_eq!(b, &[9, 8, 7, 0xFF]);
+
+        // Truncated tail: 5 bytes = one whole pixel + a 2-byte remainder that must be dropped.
+        let mut scratch = Vec::new();
+        let (_, b) = normalize_cpu_rgb(PixelFormat::Rgb, &[1, 2, 3, 4, 5], &mut scratch, false);
+        assert_eq!(b, &[1, 2, 3, 0xFF]);
+
+        // 4-bpp passthrough: borrowed, scratch untouched.
+        let src = [10u8, 20, 30, 40];
+        let mut scratch = Vec::new();
+        let (f, b) = normalize_cpu_rgb(PixelFormat::Bgrx, &src, &mut scratch, false);
+        assert_eq!(f, PixelFormat::Bgrx);
+        assert!(std::ptr::eq(b.as_ptr(), src.as_ptr()));
+        assert!(scratch.is_empty());
+
+        // The 4-bpp mapping the expand lands on matches pixel_to_vk's existing table.
+        assert_eq!(
+            pixel_to_vk(PixelFormat::Rgbx),
+            Some(vk::Format::R8G8B8A8_UNORM)
+        );
+        assert_eq!(
+            pixel_to_vk(PixelFormat::Bgrx),
+            Some(vk::Format::B8G8R8A8_UNORM)
+        );
+    }
+
+    /// RGB-direct mode (`bgra_target = true`): everything lands in B,G,R,X order because the
+    /// video session's `pictureFormat` is `B8G8R8A8_UNORM` and the encode source must match it
+    /// (VUID-vkCmdEncodeVideoKHR-pEncodeInfo-08207 — caught live on RADV). B-first inputs pass
+    /// through borrowed; R-first inputs are channel-swapped, 3-bpp and 4-bpp alike.
+    #[test]
+    fn normalize_cpu_rgb_forces_bgra_for_the_encode_source() {
+        // Rgb (R,G,B) → B,G,R,X with the swap folded into the expand.
+        let mut scratch = Vec::new();
+        let (f, b) = normalize_cpu_rgb(PixelFormat::Rgb, &[1, 2, 3], &mut scratch, true);
+        assert_eq!(f, PixelFormat::Bgrx);
+        assert_eq!(b, &[3, 2, 1, 0xFF]);
+
+        // Bgr (B,G,R) → same order, expanded.
+        let mut scratch = Vec::new();
+        let (f, b) = normalize_cpu_rgb(PixelFormat::Bgr, &[9, 8, 7], &mut scratch, true);
+        assert_eq!(f, PixelFormat::Bgrx);
+        assert_eq!(b, &[9, 8, 7, 0xFF]);
+
+        // Rgbx: 4-bpp but R-first — swapped, alpha replaced by the 0xFF pad.
+        let mut scratch = Vec::new();
+        let (f, b) = normalize_cpu_rgb(PixelFormat::Rgbx, &[1, 2, 3, 4], &mut scratch, true);
+        assert_eq!(f, PixelFormat::Bgrx);
+        assert_eq!(b, &[3, 2, 1, 0xFF]);
+
+        // Bgrx/Bgra already match the session order: borrowed untouched.
+        let src = [10u8, 20, 30, 40];
+        let mut scratch = Vec::new();
+        let (f, b) = normalize_cpu_rgb(PixelFormat::Bgra, &src, &mut scratch, true);
+        assert_eq!(f, PixelFormat::Bgra);
+        assert!(std::ptr::eq(b.as_ptr(), src.as_ptr()));
+        assert!(scratch.is_empty());
+    }
+}
@@ -80,3 +80,74 @@ pub struct VideoEncodeSessionRgbConversionCreateInfoVALVE {
 pub fn stype(raw: i32) -> vk::StructureType {
    vk::StructureType::from_raw(raw)
 }
+
+// ---------- ABI layout guard ----------
+//
+// These structs are hand-copied from the registry and handed to the driver through raw `p_next`
+// chains, so nothing in the type system relates them to the C definitions any more: an edit that
+// inserts, drops, widens or re-pads a field is not a compile error, it is the driver reading our
+// bytes at the wrong offsets. The assertions below are the missing compile error. They are `const`
+// rather than `#[cfg(test)]` (the shape `amf.rs` uses) so they hold in every build, including the
+// shipped one, and on any target this module compiles for.
+//
+// What they catch: a changed field width, an inserted or removed field, a changed array length, a
+// padding assumption that only holds on one target. What they CANNOT catch: swapping two fields of
+// the same type — offsets are unchanged. That case is only caught by reading the registry, so the
+// field order here was diffed against `vulkan_core.h` (Vulkan-Headers `main`, 2026-07-25) when
+// these assertions were written, along with every `ST_*` and flag-bit value above.
+//
+// Deliberately duplicated in `vk_av1_encode.rs` rather than shared: both modules exist to be
+// deleted wholesale once `ash` ships these bindings, and a shared helper would make deleting one
+// break the other.
+macro_rules! assert_abi_layout {
+    ($t:ty { size: $size:expr, align: $align:expr $(, $field:ident @ $off:expr)* $(,)? }) => {
+        const _: () = {
+            assert!(
+                ::core::mem::size_of::<$t>() == $size,
+                concat!(stringify!($t), ": size does not match the C ABI")
+            );
+            assert!(
+                ::core::mem::align_of::<$t>() == $align,
+                concat!(stringify!($t), ": alignment does not match the C ABI")
+            );
+            $(assert!(
+                ::core::mem::offset_of!($t, $field) == $off,
+                concat!(stringify!($t), ".", stringify!($field), ": offset does not match the C ABI")
+            );)*
+        };
+    };
+}
+
+assert_abi_layout!(PhysicalDeviceVideoEncodeRgbConversionFeaturesVALVE {
+    size: 24, align: 8,
+    s_type @ 0,
+    p_next @ 8,
+    video_encode_rgb_conversion @ 16,
+});
+
+assert_abi_layout!(VideoEncodeRgbConversionCapabilitiesVALVE {
+    size: 32, align: 8,
+    s_type @ 0,
+    p_next @ 8,
+    rgb_models @ 16,
+    rgb_ranges @ 20,
+    x_chroma_offsets @ 24,
+    y_chroma_offsets @ 28,
+});
+
+assert_abi_layout!(VideoEncodeProfileRgbConversionInfoVALVE {
+    size: 24, align: 8,
+    s_type @ 0,
+    p_next @ 8,
+    perform_encode_rgb_conversion @ 16,
+});
+
+assert_abi_layout!(VideoEncodeSessionRgbConversionCreateInfoVALVE {
+    size: 32, align: 8,
+    s_type @ 0,
+    p_next @ 8,
+    rgb_model @ 16,
+    rgb_range @ 20,
+    x_chroma_offset @ 24,
+    y_chroma_offset @ 28,
+});
@@ -67,12 +67,385 @@ pub(super) fn resolve_subframe(default_on: bool) -> bool {
    }
 }

+/// Resolved NVENC split-frame encode mode for a session — ONE selector shared by the Windows and
+/// Linux direct-SDK backends (they had drifted into byte-identical duplicates, one of which
+/// logged and one didn't). Precedence:
+/// 1. `PUNKTFUNK_SPLIT_ENCODE` = `0`/`disable` | `1`/`auto` (AUTO_FORCED) | `2` | `3` — operator
+///    override, always wins.
+/// 2. 10-bit → DISABLE: 2-way split is measurably SLOWER on Ada for Main10 — at 5120×1440@240
+///    forced-2 took 7.6 ms/frame (~131 fps) vs 2.8 ms (~357 fps) single-engine (the split/merge
+///    overhead dominates), and a single engine handles 5K@240 Main10 well under budget. This was
+///    the "broken animations in HDR" cap at ~131 fps.
+/// 3. Pixel rate ≥ [`super::SPLIT_FORCE_PIXEL_RATE`] → force 2-way (AUTO never engages below
+///    ~2112 px height, so 4K120 must be forced onto the second engine).
+/// 4. Else AUTO (the ~2% BD-rate split cost isn't worth it at low pixel rates).
+///
+/// The caller still owns the rejection fallback (retry split-disabled) — a codec/config that
+/// rejects the chosen mode downgrades at open, not here.
+pub(super) fn resolve_split_mode(bit_depth: u8, pixel_rate: u64) -> u32 {
+    use nv::NV_ENC_SPLIT_ENCODE_MODE as M;
+    let mode = match std::env::var("PUNKTFUNK_SPLIT_ENCODE").ok().as_deref() {
+        Some("0") | Some("disable") => M::NV_ENC_SPLIT_DISABLE_MODE as u32,
+        Some("1") | Some("auto") => M::NV_ENC_SPLIT_AUTO_FORCED_MODE as u32,
+        Some("3") => M::NV_ENC_SPLIT_THREE_FORCED_MODE as u32,
+        Some("2") => M::NV_ENC_SPLIT_TWO_FORCED_MODE as u32,
+        _ if bit_depth >= 10 => M::NV_ENC_SPLIT_DISABLE_MODE as u32,
+        _ if pixel_rate >= super::SPLIT_FORCE_PIXEL_RATE => M::NV_ENC_SPLIT_TWO_FORCED_MODE as u32,
+        _ => M::NV_ENC_SPLIT_AUTO_MODE as u32,
+    };
+    tracing::debug!(
+        split_mode = mode,
+        bit_depth,
+        pixel_rate,
+        "NVENC split-encode mode selected"
+    );
+    mode
+}
+
+/// One session config's identity for the process-lifetime bitrate-ceiling cache
+/// ([`cached_ceiling`]/[`store_ceiling`]). Everything the driver's codec-level validation keys
+/// off: the GPU (different NVENC generations have different level ceilings), dims/fps (the luma
+/// rate selects the level), depth/chroma (they select the profile) and the split mode the
+/// sessions ACTUALLY opened with (a split session budgets per engine).
+#[derive(Clone, Copy, PartialEq, Eq, Hash)]
+pub(super) struct CeilingKey {
+    /// GPU identity — Linux: the process-global shared `CUcontext` pointer; Windows: the render
+    /// adapter LUID (0 when unresolved). Best effort: the cache is advisory (see
+    /// [`cached_ceiling`]), so a colliding identity costs one failed open + re-search, never a
+    /// wrong session.
+    pub gpu: u64,
+    pub codec: Codec,
+    pub width: u32,
+    pub height: u32,
+    pub fps: u32,
+    pub bit_depth: u8,
+    pub chroma_444: bool,
+    pub split_mode: u32,
+}
+
+fn ceilings() -> &'static std::sync::Mutex<std::collections::HashMap<CeilingKey, u64>> {
+    static CEILINGS: std::sync::OnceLock<
+        std::sync::Mutex<std::collections::HashMap<CeilingKey, u64>>,
+    > = std::sync::OnceLock::new();
+    CEILINGS.get_or_init(Default::default)
+}
+
+/// The codec-level bitrate ceiling (bps) a previous clamp search discovered for `key` this
+/// process lifetime, if any. ADVISORY: the consumer must treat a failed open at the cached value
+/// as a stale entry (fall back to the full search, which rewrites it via [`store_ceiling`]) —
+/// that self-healing is what lets the key's GPU identity be best-effort. What this buys: an ABR
+/// overshoot on a config whose ceiling is already known opens (or in-place reconfigures) straight
+/// AT the ceiling instead of re-running the ~6-open binary search and its ~half-second of session
+/// churn per rebuild.
+pub(super) fn cached_ceiling(key: &CeilingKey) -> Option<u64> {
+    ceilings().lock().unwrap().get(key).copied()
+}
+
+/// Record the clamp search's discovered max accepted bitrate (bps) for `key`.
+pub(super) fn store_ceiling(key: CeilingKey, bps: u64) {
+    ceilings().lock().unwrap().insert(key, bps);
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use nv::NV_ENC_SPLIT_ENCODE_MODE as M;
+
+    // These assume PUNKTFUNK_SPLIT_ENCODE is unset (CI); an operator override deliberately wins.
+
+    /// `encodeCodecConfig` is a C union, so the HEVC 4:4:4 arm must be codec-gated or it stamps
+    /// `hevcConfig` bytes onto another codec's config. Before the gate this branch was reached on
+    /// ANY codec with `chroma_444 && full_chroma_input` and stayed non-UB only because `lib.rs`
+    /// degrades 4:4:4 for non-HEVC — a two-file invariant with nothing asserting it.
+    ///
+    /// It also had to stop swallowing the per-codec bit-depth arm: this is an `if`/`else if`, so a
+    /// non-HEVC 4:4:4 session used to take the HEVC branch and get NEITHER 4:4:4 nor its own 10-bit
+    /// setup. AV1 asserts the depth it actually needs.
+    fn low_latency_cfg(codec: Codec, chroma_444: bool, bit_depth: u8) -> LowLatencyConfig {
+        LowLatencyConfig {
+            codec,
+            bitrate: 20_000_000,
+            fps: 60,
+            custom_vbv: false,
+            chroma_444,
+            full_chroma_input: true,
+            bit_depth,
+            av1_input_depth_minus8: 0,
+            hdr: false,
+            rfi_supported: false,
+            slices: 0,
+        }
+    }
+
+    #[test]
+    fn hevc_444_still_takes_the_frext_path() {
+        // `NV_ENC_CONFIG` must NOT be `mem::zeroed` — `frameFieldMode`/`mvPrecision` are C enums
+        // whose discriminants start at 1, so all-zero is not a valid value and Rust's own
+        // zero-init check aborts the process. Production seeds it the same way, from `Default`
+        // (then overwrites from the driver's preset).
+        // SAFETY: `apply_low_latency_config` only writes into the caller's config (union writes
+        // included) and makes no driver calls, so this is pure in-memory work.
+        let cfg = unsafe {
+            let mut cfg = nv::NV_ENC_CONFIG {
+                version: nv::NV_ENC_CONFIG_VER,
+                ..Default::default()
+            };
+            apply_low_latency_config(&mut cfg, low_latency_cfg(Codec::H265, true, 10));
+            cfg
+        };
+        assert_eq!(cfg.profileGUID, nv::NV_ENC_HEVC_PROFILE_FREXT_GUID);
+        // SAFETY: an HEVC session's union arm is `hevcConfig` — the one this path wrote.
+        unsafe {
+            assert_eq!(cfg.encodeCodecConfig.hevcConfig.chromaFormatIDC(), 3);
+            assert_eq!(cfg.encodeCodecConfig.hevcConfig.pixelBitDepthMinus8(), 2);
+        }
+    }
+
+    #[test]
+    fn av1_never_takes_the_hevc_444_union_write() {
+        // SAFETY: as above — pure in-memory config authoring, no driver involvement.
+        let cfg = unsafe {
+            let mut cfg = nv::NV_ENC_CONFIG {
+                version: nv::NV_ENC_CONFIG_VER,
+                ..Default::default()
+            };
+            apply_low_latency_config(&mut cfg, low_latency_cfg(Codec::Av1, true, 10));
+            cfg
+        };
+        // The HEVC FREXT profile GUID on an AV1 session is an INVALID_PARAM at open.
+        assert_ne!(
+            cfg.profileGUID,
+            nv::NV_ENC_HEVC_PROFILE_FREXT_GUID,
+            "4:4:4 on AV1 must not stamp the HEVC FREXT profile"
+        );
+        // ...and the AV1 arm must still have run, which the old if/else-if skipped entirely.
+        // SAFETY: an AV1 session's union arm is `av1Config`.
+        unsafe {
+            assert_eq!(
+                cfg.encodeCodecConfig.av1Config.pixelBitDepthMinus8(),
+                2,
+                "AV1 10-bit setup was swallowed by the HEVC 4:4:4 branch"
+            );
+        }
+    }
+
+    #[test]
+    fn split_forces_two_way_at_4k120() {
+        // The regression this threshold constant exists for: 3840×2160×120 = 995,328,000 sat
+        // 0.47% under the old `> 1_000_000_000` gate and stayed AUTO — pinned ~107 fps on a
+        // 4090 because AUTO never engages at 2160 px height.
+        let four_k_120 = 3840u64 * 2160 * 120;
+        assert_eq!(
+            resolve_split_mode(8, four_k_120),
+            M::NV_ENC_SPLIT_TWO_FORCED_MODE as u32
+        );
+    }
+
+    #[test]
+    fn split_leaves_1440p240_auto() {
+        // 884.7 Mpix/s is comfortably single-engine — the threshold move must not drag it in.
+        let qhd_240 = 2560u64 * 1440 * 240;
+        assert_eq!(
+            resolve_split_mode(8, qhd_240),
+            M::NV_ENC_SPLIT_AUTO_MODE as u32
+        );
+    }
+
+    #[test]
+    fn split_disabled_for_10bit_even_at_high_pixel_rate() {
+        // The measured Main10 rule: split/merge overhead dominates 10-bit on Ada (7.6 ms forced-2
+        // vs 2.8 ms single-engine at 5K240) — 10-bit precedes the pixel-rate arm.
+        let five_k_240 = 5120u64 * 1440 * 240;
+        assert_eq!(
+            resolve_split_mode(10, five_k_240),
+            M::NV_ENC_SPLIT_DISABLE_MODE as u32
+        );
+    }
+
+    #[test]
+    fn ceiling_cache_round_trips_and_keys_precisely() {
+        let key = CeilingKey {
+            gpu: 0xB0B0,
+            codec: Codec::H265,
+            width: 3840,
+            height: 2160,
+            fps: 120,
+            bit_depth: 8,
+            chroma_444: false,
+            split_mode: M::NV_ENC_SPLIT_TWO_FORCED_MODE as u32,
+        };
+        assert_eq!(cached_ceiling(&key), None);
+        store_ceiling(key, 794_000_000);
+        assert_eq!(cached_ceiling(&key), Some(794_000_000));
+        // Any config-identity change is a different ceiling — a miss, never a wrong clamp.
+        assert_eq!(cached_ceiling(&CeilingKey { fps: 60, ..key }), None);
+        assert_eq!(
+            cached_ceiling(&CeilingKey {
+                split_mode: M::NV_ENC_SPLIT_DISABLE_MODE as u32,
+                ..key
+            }),
+            None
+        );
+        // A re-search overwrites (the advisory-cache stale-entry path).
+        store_ceiling(key, 620_000_000);
+        assert_eq!(cached_ceiling(&key), Some(620_000_000));
+    }
+}
+
 /// Reference-frame DPB depth when RFI is supported (Apollo uses 5). A deeper DPB lets an invalidated
 /// reference fall back to an older still-valid frame instead of a full IDR; `numRefL0 = 1` keeps each
-/// P-frame single-reference for low latency. Also the window the backends' `invalidate_ref_frames`
-/// paths check against (`frame_idx - RFI_DPB` = the oldest frame still in the DPB).
+/// P-frame single-reference for low latency. Also the window [`plan_range_recovery`] checks against
+/// (`next_ts - RFI_DPB` = the oldest frame still in the DPB).
 pub(super) const RFI_DPB: u32 = 5;

+/// One loss event's recovery decision for the timestamp-range RFI both direct-NVENC backends run
+/// (the range half of WP7.2's policy extraction; the slot half — AMF/QSV/Vulkan — is
+/// `crate::rfi`). The mechanism (the per-timestamp `nvEncInvalidateRefFrames` loop, the
+/// `last_rfi_range`/`pending_anchor` stores, the null-handle/`rfi_supported` gate) stays in each
+/// backend.
+pub(super) enum RangePlan {
+    /// The last successful invalidation already covers this range — no new driver calls, no IDR.
+    /// The caller must still RE-ARM its recovery anchor: the client re-asking means the previous
+    /// anchor AU may itself have been lost, and the next frame is just as clean a re-anchor.
+    Covered,
+    /// Invalidate `first..=last` (the CLAMPED range — this is also what the caller must record in
+    /// `last_rfi_range` on success, exactly as the inline code stored the post-clamp values).
+    Invalidate { first: i64, last: i64 },
+    /// Recovery without an IDR is impossible (nonsense range, loss older than the DPB, or a range
+    /// entirely in the future) — the caller returns `false` and its (coalesced) keyframe path
+    /// recovers. Deliberately NOT paired with any state clearing: neither twin touches
+    /// `pending_anchor` on decline (matching Vulkan's decline, opposite of AMF/QSV's
+    /// `pending_force` clear — see `crate::rfi`'s module doc before "harmonizing").
+    Decline,
+}
+
+/// The range-RFI policy, extracted verbatim from the two backends' `invalidate_ref_frames` (they
+/// were hand-copied twins). Step order is load-bearing and pinned by tests:
+///
+/// 1. nonsense range (`first < 0 || first > last`) → [`RangePlan::Decline`];
+/// 2. covering-range dedup — checked with the UNCLAMPED `last`, BEFORE the DPB window, so a
+///    covered re-ask never touches the driver even when the range has since left the DPB;
+/// 3. DPB window: `first < next_ts - RFI_DPB` → Decline (a lost frame older than the DPB cannot
+///    be invalidated; the only correct recovery is an IDR);
+/// 4. clamp `last` to `next_ts - 1` (never invalidate a timestamp never assigned); an inverted
+///    range after the clamp (loss entirely in the future — a prediction desync) → Decline.
+///
+/// `next_ts` is the backend's `frame_idx`: the NEXT timestamp to assign, which `submit_indexed`
+/// pins to the wire frame index — so the client's lost-frame range maps 1:1 onto the timestamps
+/// the driver invalidates, across every rebuild/reset. Note `teardown()` clears `last_rfi_range`
+/// but NOT `frame_idx`, so a post-reset call legitimately sees a stale-high `next_ts` with a
+/// `None` range — the same view the inline code had.
+pub(super) fn plan_range_recovery(
+    first: i64,
+    last: i64,
+    next_ts: i64,
+    last_rfi_range: Option<(i64, i64)>,
+) -> RangePlan {
+    if first < 0 || first > last {
+        return RangePlan::Decline;
+    }
+    if let Some((pf, pl)) = last_rfi_range {
+        if first >= pf && last <= pl {
+            return RangePlan::Covered;
+        }
+    }
+    let oldest_in_dpb = next_ts - RFI_DPB as i64;
+    if first < oldest_in_dpb {
+        return RangePlan::Decline;
+    }
+    let last = last.min(next_ts - 1);
+    if first > last {
+        return RangePlan::Decline;
+    }
+    RangePlan::Invalidate { first, last }
+}
+
+#[cfg(test)]
+mod range_policy_tests {
+    use super::{plan_range_recovery, RangePlan, RFI_DPB};
+
+    /// Convenience: the plan with no prior invalidation recorded.
+    fn plan(first: i64, last: i64, next_ts: i64) -> RangePlan {
+        plan_range_recovery(first, last, next_ts, None)
+    }
+
+    #[test]
+    fn nonsense_ranges_decline() {
+        assert!(matches!(plan(-1, 5, 100), RangePlan::Decline));
+        assert!(matches!(plan(7, 5, 100), RangePlan::Decline));
+    }
+
+    #[test]
+    fn covering_range_dedups_partial_overlap_does_not() {
+        let prior = Some((90i64, 95i64));
+        // Exact cover and sub-range → Covered. This pins EXISTING behavior, including that a
+        // covered range survives a forced IDR with zero driver calls (nothing clears
+        // `last_rfi_range` on a keyframe) — a recorded fact, not an endorsement.
+        assert!(matches!(
+            plan_range_recovery(90, 95, 100, prior),
+            RangePlan::Covered
+        ));
+        assert!(matches!(
+            plan_range_recovery(92, 94, 100, prior),
+            RangePlan::Covered
+        ));
+        // Partial overlap re-invalidates the FULL new range (next_ts = 98 keeps the window open:
+        // oldest_in_dpb = 93; at next_ts = 100 the same range would age out and Decline instead).
+        assert!(matches!(
+            plan_range_recovery(93, 97, 98, prior),
+            RangePlan::Invalidate {
+                first: 93,
+                last: 97
+            }
+        ));
+    }
+
+    /// The covering check runs BEFORE the DPB window: a covered re-ask stays Covered (no driver
+    /// calls needed) even when the range has since aged out of the DPB.
+    #[test]
+    fn covered_is_checked_before_the_dpb_window() {
+        let prior = Some((10i64, 12i64));
+        assert!(matches!(
+            plan_range_recovery(10, 12, 100, prior),
+            RangePlan::Covered
+        ));
+        // ...whereas the same range with no prior invalidation is outside the window → Decline.
+        assert!(matches!(plan(10, 12, 100), RangePlan::Decline));
+    }
+
+    #[test]
+    fn dpb_window_boundary() {
+        let next_ts = 100i64;
+        let oldest = next_ts - RFI_DPB as i64; // 95: the oldest timestamp still in the DPB
+        assert!(matches!(
+            plan(oldest, oldest, next_ts),
+            RangePlan::Invalidate { .. }
+        ));
+        assert!(matches!(
+            plan(oldest - 1, oldest, next_ts),
+            RangePlan::Decline
+        ));
+    }
+
+    /// `last` clamps to `next_ts - 1` (the newest encoded frame); the Invalidate carries the
+    /// CLAMPED value — which is also what the caller records in `last_rfi_range`.
+    #[test]
+    fn clamps_to_newest_encoded() {
+        assert!(matches!(
+            plan(98, 150, 100),
+            RangePlan::Invalidate {
+                first: 98,
+                last: 99
+            }
+        ));
+        // A range entirely in the future inverts under the clamp → Decline (prediction desync).
+        assert!(matches!(plan(100, 150, 100), RangePlan::Decline));
+        // Fresh session (`frame_idx == 0`): window passes (oldest = -5) but the clamp gives
+        // last = -1 < first → Decline. The inline code behaved identically.
+        assert!(matches!(plan(0, 3, 0), RangePlan::Decline));
+    }
+}
+
 /// The per-session knobs both direct-NVENC backends feed [`apply_low_latency_config`]. `Copy` so the
 /// backend fills it from `self` at the call. The two input-format fields bridge the only real
 /// divergence between the CUDA and D3D11 paths (which surface formats can carry full chroma / 10-bit
@@ -219,7 +592,27 @@ pub(super) unsafe fn apply_low_latency_config(cfg: &mut nv::NV_ENC_CONFIG, c: Lo
    // profile) takes precedence and composes with 10-bit (Main 4:4:4 10); it needs a full-chroma-
    // capable input. Otherwise 10-bit selects Main10 (HEVC) or the AV1 output depth — stamping the
    // HEVC Main10 GUID onto an AV1 session is an INVALID_PARAM, so bit depth is set PER CODEC.
-    if c.chroma_444 && c.full_chroma_input {
+    // `encodeCodecConfig` is a C UNION, so the `hevcConfig` writes below are only meaningful on an
+    // HEVC session — on an H.264 or AV1 one they reinterpret that codec's own config bytes. The
+    // codec test is therefore load-bearing, not defensive: without it this branch was gated purely
+    // on `chroma_444 && full_chroma_input` and stayed non-UB only because `lib.rs` degrades 4:4:4
+    // for non-HEVC codecs. That was a two-file invariant with nothing asserting it, on the path
+    // BOTH direct-NVENC backends take.
+    //
+    // Being a codec test also fixes a second, quieter bug in the same shape: this is an
+    // `if`/`else if`, so a non-HEVC session that somehow arrived with `chroma_444` set took this
+    // branch and skipped the per-codec bit-depth arm entirely — ending up with neither HEVC 4:4:4
+    // (wrong for it) nor its own 10-bit configuration (simply missing). Non-HEVC now falls through
+    // to the arm that knows what to do with it.
+    let want_444 = c.chroma_444 && c.full_chroma_input;
+    if want_444 && c.codec != Codec::H265 {
+        tracing::warn!(
+            codec = ?c.codec,
+            "4:4:4 requested on a non-HEVC NVENC session — ignoring it (Range Extensions are \
+             HEVC-only); the negotiator should have degraded this to 4:2:0 before the open"
+        );
+    }
+    if want_444 && c.codec == Codec::H265 {
        cfg.profileGUID = nv::NV_ENC_HEVC_PROFILE_FREXT_GUID;
        cfg.encodeCodecConfig.hevcConfig.set_chromaFormatIDC(3);
        if c.bit_depth == 10 {
@@ -8,27 +8,77 @@
 //! means and what the operator should do, and folds that cause into the `anyhow::Error` at
 //! construction, so every downstream `{e:#}` log (the encode-recovery loop, session teardown) says
 //! the useful thing without extra plumbing.
+//!
+//! One status needs process state to explain honestly: the driver reports BOTH "your headers are
+//! newer than my kernel module" and "I can no longer hand this process a session" as
+//! `NV_ENC_ERR_INVALID_VERSION`. [`note_session_opened`] latches the fact that a session already
+//! opened here, which tells the two apart — see [`explain`].
+
+use std::sync::atomic::{AtomicBool, Ordering};

 use nvidia_video_codec_sdk::sys::nvEncodeAPI as nv;

+/// Latched the first time `nvEncOpenEncodeSessionEx` succeeds in this process (the caps probe, a
+/// real session open, or the Windows availability probe — every one of them completes the
+/// userspace↔kernel-module handshake).
+///
+/// The load-time gate (`NvEncodeAPIGetMaxSupportedVersion`, both backends' `load_api`) can NOT
+/// serve this purpose: it is a pure userspace query, so the genuine "updated the driver, didn't
+/// reboot" skew sails through it and fails later, at the open. Only a session that actually opened
+/// proves the kernel module agreed.
+static SESSION_OPENED: AtomicBool = AtomicBool::new(false);
+
+/// Record that an NVENC session opened. Call right after every successful
+/// `open_encode_session_ex`, so [`explain`] can rule a version skew out for the rest of the
+/// process.
+pub(super) fn note_session_opened() {
+    SESSION_OPENED.store(true, Ordering::Relaxed);
+}
+
+/// The two very different failures the driver reports as `NV_ENC_ERR_INVALID_VERSION`, split on
+/// whether a session has already opened here (`session_opened`). Pure, so both halves are testable
+/// without touching the process-wide latch.
+fn invalid_version(session_opened: bool) -> String {
+    if session_opened {
+        // Same status, opposite cause: a session ALREADY opened in this process, so the driver's
+        // kernel module accepted this exact build's version word minutes ago. A version skew is
+        // static — it cannot come and go — so "update the driver / reboot" is the wrong advice
+        // here, and following it costs the operator a reboot per stream (2026-07 field report: one
+        // stream works, the next fails at the caps probe, forever, until the PROCESS restarts).
+        // What is left is per-process driver state: a resource the last session did not give back,
+        // or a wedged/lost device. Say that, and point at the cheap fix.
+        // Worded for ANY call (`explain` also serves `lock_bitstream`); `call_err` already names
+        // the entry point ahead of this text, so it must not assume the session open.
+        return "this process already opened an NVENC session successfully, so this is NOT a driver \
+                version mismatch — that cannot come and go within a process, and a reboot is not \
+                the fix. The NVIDIA driver state in THIS process is exhausted or wedged: restart \
+                the Punktfunk host service to clear it, and please report this with the host log \
+                so it can be fixed properly"
+            .to_string();
+    }
+    // No session has ever opened here, so the version word really is in question. Either the
+    // driver is genuinely older than our headers, or (the sneaky case) the userspace
+    // `libnvidia-encode` reports a new-enough version to the pre-flight probe but the running
+    // kernel module is older and rejects the session — the classic "updated the driver, didn't
+    // reboot" skew. Both heal the same way.
+    format!(
+        "the NVIDIA driver is older than this build's NVENC headers (needs NVENC API {}.{} or \
+         newer), or the userspace and kernel-module driver versions are mismatched — common right \
+         after a driver update without a reboot. Update the NVIDIA driver, or reboot if you just \
+         updated it (a host restart is the usual fix).",
+        nv::NVENCAPI_MAJOR_VERSION,
+        nv::NVENCAPI_MINOR_VERSION,
+    )
+}
+
 /// A one-line, operator-actionable cause for an NVENC status. Does not repeat the raw code —
 /// callers print that alongside (see [`call_err`]). Public for the few sites that build a
 /// `String`/`format!` error instead of an `anyhow::Error`.
 pub(super) fn explain(status: nv::NVENCSTATUS) -> String {
    match status {
-        // The one this whole module exists for: a version word the driver rejects. Either the
-        // driver is genuinely older than our headers, or (the sneaky case) the userspace
-        // `libnvidia-encode` reports a new-enough version to the pre-flight probe but the running
-        // kernel module is older and rejects the session — the classic "updated the driver, didn't
-        // reboot" skew. Both heal the same way.
-        nv::NVENCSTATUS::NV_ENC_ERR_INVALID_VERSION => format!(
-            "the NVIDIA driver is older than this build's NVENC headers (needs NVENC API {}.{} or \
-             newer), or the userspace and kernel-module driver versions are mismatched — common \
-             right after a driver update without a reboot. Update the NVIDIA driver, or reboot if \
-             you just updated it (a host restart is the usual fix).",
-            nv::NVENCAPI_MAJOR_VERSION,
-            nv::NVENCAPI_MINOR_VERSION,
-        ),
+        nv::NVENCSTATUS::NV_ENC_ERR_INVALID_VERSION => {
+            invalid_version(SESSION_OPENED.load(Ordering::Relaxed))
+        }
        nv::NVENCSTATUS::NV_ENC_ERR_NO_ENCODE_DEVICE => {
            "this GPU exposes no usable NVENC engine — it has no hardware video encoder, or NVENC is \
             disabled on this card"
@@ -72,9 +122,144 @@ pub(super) fn explain(status: nv::NVENCSTATUS) -> String {
    }
 }

+/// Typed root of a failed NVENC entry-point call: carries the raw status so callers can classify
+/// the failure class, not just print it — the bitrate-clamp search must only read a
+/// parameter/caps rejection as "above the codec-level ceiling"; a transient failure shrinking the
+/// search would discover (and cache) a bogus ceiling. Recover it through an `anyhow` chain with
+/// `err.downcast_ref::<NvCallError>()` (see [`is_param_rejection`]).
+#[derive(Debug)]
+pub(super) struct NvCallError(pub(super) nv::NVENCSTATUS);
+
+impl std::fmt::Display for NvCallError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{:?} — {}", self.0, explain(self.0))
+    }
+}
+
+impl std::error::Error for NvCallError {}
+
+/// Whether `err` is an NVENC parameter/capability rejection: the driver understood the request
+/// and says THIS config is not encodable — the clamp search's "bitrate above the ceiling"
+/// evidence. Everything else (busy engine, session limit, OOM, device loss, version skew) is
+/// environmental and must propagate instead of steering the search.
+pub(super) fn is_param_rejection(err: &anyhow::Error) -> bool {
+    matches!(
+        err.downcast_ref::<NvCallError>(),
+        Some(NvCallError(
+            nv::NVENCSTATUS::NV_ENC_ERR_INVALID_PARAM
+                | nv::NVENCSTATUS::NV_ENC_ERR_UNSUPPORTED_PARAM
+                | nv::NVENCSTATUS::NV_ENC_ERR_UNIMPLEMENTED,
+        ))
+    )
+}
+
 /// Build an actionable `anyhow::Error` for a failed NVENC entry-point call. `call` names the API
-/// (e.g. `"open_encode_session_ex"`); the message carries both the raw status and its real-world
-/// cause, so triage never again reads a version mismatch as "(no NVIDIA GPU?)".
+/// (e.g. `"open_encode_session_ex"`); the chain carries both the raw status and its real-world
+/// cause, so triage never again reads a version mismatch as "(no NVIDIA GPU?)". The
+/// [`NvCallError`] root keeps the status downcastable for failure-class checks.
 pub(super) fn call_err(call: &str, status: nv::NVENCSTATUS) -> anyhow::Error {
-    anyhow::anyhow!("NVENC {call} failed: {status:?} — {}", explain(status))
+    anyhow::Error::new(NvCallError(status)).context(format!("NVENC {call} failed"))
+}
+
+/// Whether a FAILED `nvEncDestroyEncoder` status PROVES the driver holds no session for the
+/// handle — i.e. the per-process concurrent-session slot is not consumed, so the session's budget
+/// units can be refunded immediately. These are the statuses the driver returns when the session
+/// or its device no longer exists on its side (a TDR/device removal reclaims every session with
+/// the context). Everything else — `GENERIC`, `ENCODER_BUSY`, OOM, ... — is AMBIGUOUS: the slot
+/// may genuinely still be held, so the caller must park the handle fail-closed (units stay
+/// charged) and let a later retry-destroy produce the proof. Splitting on proof is what keeps the
+/// session budget from drifting low on failures (over-admitting parallel displays) WITHOUT letting
+/// one transient wedge episode permanently poison admission until a host restart.
+///
+/// Used by the Windows D3D11 backend's teardown accounting; the Linux CUDA backend has no session
+/// budget (parallel-display admission is a Windows feature), so there this exists for the unit
+/// tests only.
+#[cfg_attr(not(target_os = "windows"), allow(dead_code))]
+pub(super) fn destroy_proves_no_session(status: nv::NVENCSTATUS) -> bool {
+    matches!(
+        status,
+        nv::NVENCSTATUS::NV_ENC_ERR_DEVICE_NOT_EXIST
+            | nv::NVENCSTATUS::NV_ENC_ERR_INVALID_ENCODERDEVICE
+            | nv::NVENCSTATUS::NV_ENC_ERR_INVALID_PTR
+            | nv::NVENCSTATUS::NV_ENC_ERR_ENCODER_NOT_INITIALIZED
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Before any session has opened, the version word IS in question — keep the skew advice.
+    #[test]
+    fn invalid_version_before_any_session_blames_the_driver_version() {
+        let msg = invalid_version(false);
+        assert!(
+            msg.contains("older than this build's NVENC headers"),
+            "{msg}"
+        );
+        assert!(msg.contains("reboot if you just updated it"), "{msg}");
+    }
+
+    /// Once a session has opened here, a skew is impossible — the message must stop sending
+    /// operators to reboot (2026-07 field report: one stream per boot, forever).
+    #[test]
+    fn invalid_version_after_a_session_blames_process_state_not_the_driver() {
+        let msg = invalid_version(true);
+        assert!(msg.contains("NOT a driver version mismatch"), "{msg}");
+        assert!(msg.contains("restart the Punktfunk host service"), "{msg}");
+        assert!(
+            !msg.contains("older than this build's NVENC headers"),
+            "must not repeat the version-skew advice: {msg}"
+        );
+        assert!(
+            !msg.contains("Update the NVIDIA driver"),
+            "must not tell the operator to update a driver that just worked: {msg}"
+        );
+    }
+
+    /// The latch is one-way and only touches this status.
+    #[test]
+    fn note_session_opened_latches() {
+        note_session_opened();
+        assert!(SESSION_OPENED.load(Ordering::Relaxed));
+        note_session_opened();
+        assert!(SESSION_OPENED.load(Ordering::Relaxed));
+        assert_eq!(
+            explain(nv::NVENCSTATUS::NV_ENC_ERR_OUT_OF_MEMORY),
+            "the GPU is out of memory"
+        );
+    }
+
+    /// Destroy-failure classification: session-gone statuses refund; everything ambiguous parks.
+    /// The split is the load-bearing part of the session-budget accounting — a wrong `true`
+    /// under-counts (over-admits parallel displays), a wrong `false` merely defers the refund to
+    /// a reap retry.
+    #[test]
+    fn destroy_classification_refunds_only_on_proof() {
+        for gone in [
+            nv::NVENCSTATUS::NV_ENC_ERR_DEVICE_NOT_EXIST,
+            nv::NVENCSTATUS::NV_ENC_ERR_INVALID_ENCODERDEVICE,
+            nv::NVENCSTATUS::NV_ENC_ERR_INVALID_PTR,
+            nv::NVENCSTATUS::NV_ENC_ERR_ENCODER_NOT_INITIALIZED,
+        ] {
+            assert!(
+                destroy_proves_no_session(gone),
+                "{gone:?} proves no session"
+            );
+        }
+        for ambiguous in [
+            nv::NVENCSTATUS::NV_ENC_ERR_GENERIC,
+            nv::NVENCSTATUS::NV_ENC_ERR_ENCODER_BUSY,
+            nv::NVENCSTATUS::NV_ENC_ERR_OUT_OF_MEMORY,
+            nv::NVENCSTATUS::NV_ENC_ERR_INVALID_PARAM,
+            // INVALID_DEVICE sounds like the gone class but is also what a transiently-confused
+            // driver returns — deliberately fail-closed (park + retry), not refunded.
+            nv::NVENCSTATUS::NV_ENC_ERR_INVALID_DEVICE,
+        ] {
+            assert!(
+                !destroy_proves_no_session(ambiguous),
+                "{ambiguous:?} must park fail-closed"
+            );
+        }
+    }
 }
@@ -246,6 +246,17 @@ mod tests {
        assert!(block_count_32x32(3840, 2160, true) <= u16::MAX as u32);
        assert!(block_count_32x32(7680, 4320, true) > u16::MAX as u32);
        assert!(block_count_32x32(7680, 4320, false) <= u16::MAX as u32);
+        // …and 4:2:0 wraps it too, just later — the hole the old 4:4:4-only open guard left.
+        // `Codec::max_dimension()` allows PyroWave 8192px per axis, so these modes were
+        // reachable from a client-requested `mode=WxHxFPS`, and the negotiator's 4:4:4 → 4:2:0
+        // downgrade routed oversized modes straight into the unguarded branch.
+        // `validate_dimensions` now rejects them against this 4:2:0 count.
+        assert_eq!(block_count_32x32(8192, 6144, false), 73728);
+        assert_eq!(block_count_32x32(8192, 8192, false), 98304);
+        assert!(block_count_32x32(8192, 6144, false) > u16::MAX as u32);
+        assert!(block_count_32x32(8192, 8192, false) > u16::MAX as u32);
+        // The largest 4:2:0 mode that still fits, for the boundary the validator enforces.
+        assert!(block_count_32x32(7680, 4320, false) <= u16::MAX as u32);
    }

    #[test]
@@ -0,0 +1,178 @@
+//! The slot-family RFI (reference-frame invalidation) recovery **policy**, shared by the three
+//! backends that answer a loss with a re-reference to a known-good older frame instead of an IDR:
+//! native AMF (user-LTR bitfield), native QSV (`mfxExtRefListCtrl` LTR) and Vulkan Video (the
+//! app-owned DPB slot table). One decision, three mechanisms — the policy lived as three
+//! hand-copies and had already diverged once (the fecbec2d taint sweep reached AMF/QSV a commit
+//! before the Vulkan backend was carved out, and Vulkan shipped without it until a later fix).
+//! The NVENC twins' *range* policy is the other half of WP7.2, in
+//! [`super::nvenc_core::plan_range_recovery`].
+//!
+//! Policy only. Every mechanism — how a force is applied, how distrust is *persisted* (AMF clears
+//! its mirror slot to `None`, QSV sets a separate `ltr_tainted` flag because its mirror must keep
+//! naming the slot for the RejectedRefList, Vulkan blanks `slot_wire` to `-1` while `slot_poc`
+//! keeps the picture resident for the RPS) — stays in its backend. Callers feed their
+//! **currently-trusted** references and apply the returned taints through their own marker; that
+//! caller-side filter is exactly what makes the three persistence schemes equivalent under one
+//! pure function.
+//!
+//! The decline arm is also the backend's: AMF/QSV clear an un-consumed `pending_force` (the sweep
+//! may have emptied the slot it points at), while Vulkan deliberately leaves `pending_loss` armed
+//! (a stale arm is re-resolved at frame-build, where a failed re-pick forces the IDR that heals
+//! the stream). Do not harmonize them here.
+
+/// One loss event's recovery decision over a slot table: which trusted references become
+/// untrustworthy, and which one anchors the recovery.
+pub(super) struct SlotPlan {
+    /// Bitmask of slots whose reference was encoded **at or after** the loss start — inside the
+    /// client's corrupt window, so the client either never received it or decoded it against a
+    /// broken chain. Serving one as "known-good" on a LATER loss ships corruption tagged as the
+    /// recovery anchor; the backend must record the distrust in its own persistent marker, because
+    /// these wires would otherwise look like valid pre-loss anchors to the next loss event.
+    pub(super) tainted: u32,
+    /// The recovery anchor: the newest trusted reference **strictly older** than the loss —
+    /// `(slot, wire)` — i.e. the most recent picture the client still holds intact, so
+    /// re-referencing it costs the smallest residual. `None`: every candidate is inside or after
+    /// the corrupt window — the caller declines and its (coalesced) keyframe path recovers.
+    pub(super) anchor: Option<(usize, i64)>,
+}
+
+/// Plan the recovery for a loss starting at wire frame `loss_first`, over the backend's
+/// currently-trusted references (`(slot, wire)`; previously-distrusted entries must already be
+/// filtered out by the caller — see the module doc). Sweep and pick are one call so the anchor is
+/// chosen from the same snapshot the taints are computed from, by construction: the anchor
+/// delegates to [`pick_anchor`], and `wire >= loss_first` (taint) and `wire < loss_first`
+/// (anchor) are disjoint, so a slot tainted by this call can never be this call's anchor.
+pub(super) fn plan_slot_recovery(refs: &[(usize, i64)], loss_first: i64) -> SlotPlan {
+    // The callers' validity gate (`first < 0 → decline`) is what makes their sentinel filters
+    // (-1 / `None`) exact views of "trusted"; this assert keeps the contract visible from inside
+    // the extracted code. Plain assert: the lint legs run --release, and a compiled-out check
+    // here would silently drop taints instead of failing loudly.
+    assert!(
+        loss_first >= 0,
+        "loss_first must be validity-gated by the caller"
+    );
+    let mut tainted = 0u32;
+    for &(slot, wire) in refs {
+        if wire >= loss_first {
+            assert!(slot < 32, "slot table exceeds the u32 taint mask");
+            tainted |= 1 << slot;
+        }
+    }
+    SlotPlan {
+        tainted,
+        anchor: pick_anchor(refs, loss_first),
+    }
+}
+
+/// The pick half alone: newest trusted reference strictly older than the loss. Ties break to the
+/// first entry in `refs` (every caller feeds ascending slot order, so the lowest slot wins —
+/// the strict `>` all three backends used). Standalone because Vulkan re-picks at frame-build
+/// time: its arm carries the loss start, not the slot, so the slot is resolved against the table
+/// as it stands when the recovery frame is actually encoded.
+pub(super) fn pick_anchor(refs: &[(usize, i64)], loss_first: i64) -> Option<(usize, i64)> {
+    let mut best: Option<(usize, i64)> = None;
+    for &(slot, wire) in refs {
+        if wire < loss_first && best.is_none_or(|(_, b)| wire > b) {
+            best = Some((slot, wire));
+        }
+    }
+    best
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{pick_anchor, plan_slot_recovery};
+
+    /// Adapt a raw slot table (the Vulkan `slot_wire` shape: `-1` = empty) into the trusted view
+    /// the policy takes — the same filter the backend adapters apply.
+    fn view(wires: &[i64]) -> Vec<(usize, i64)> {
+        wires
+            .iter()
+            .enumerate()
+            .filter_map(|(s, &w)| (w >= 0).then_some((s, w)))
+            .collect()
+    }
+
+    /// Apply a plan's taints the way the Vulkan adapter does (blank the wire) — the persistence
+    /// half the pure fn hands back to the caller.
+    fn apply(wires: &mut [i64], tainted: u32) {
+        for (s, w) in wires.iter_mut().enumerate() {
+            if tainted & (1 << s) != 0 {
+                *w = -1;
+            }
+        }
+    }
+
+    /// The RFI anchor picker: newest resident wire strictly older than the loss; empty/newer
+    /// slots never qualify. (Migrated 1:1 from `vulkan_video.rs`'s `pick_recovery_slot` tests —
+    /// same vectors, now `(slot, wire)`-valued.)
+    #[test]
+    fn picks_newest_pre_loss() {
+        // slots hold wires 5..12 (ring position arbitrary); loss starts at 9 → anchor = wire 8.
+        let wires = [8i64, 9, 10, 11, 12, 5, 6, 7];
+        assert_eq!(pick_anchor(&view(&wires), 9), Some((0, 8)));
+        // loss older than everything resident → no anchor (caller keyframes).
+        assert_eq!(pick_anchor(&view(&wires), 5), None);
+        // empty slots (-1) are skipped by the view filter and never anchor.
+        assert_eq!(pick_anchor(&view(&[-1, 3, -1, 4]), 5), Some((3, 4)));
+        assert_eq!(pick_anchor(&view(&[-1; 8]), 5), None);
+        // wire == loss_first is INSIDE the corrupt window: strictly-older only.
+        assert_eq!(pick_anchor(&view(&[9, 8]), 9), Some((1, 8)));
+        // tie on the wire → first entry (= lowest slot) wins, the strict `>` all three backends
+        // used.
+        assert_eq!(pick_anchor(&[(2, 7), (5, 7)], 9), Some((2, 7)));
+        // empty view.
+        assert_eq!(pick_anchor(&[], 9), None);
+    }
+
+    /// The taint sweep (fecbec2d's fix): a slot encoded inside an EARLIER, still unrepaired loss
+    /// window must not become the "known-good" anchor of a LATER loss. Without persisted
+    /// distrust the picker accepts it — it is resident and its wire is below the second loss
+    /// start — and the frame ships tagged `recovery_anchor`, lifting the client's freeze onto a
+    /// reference it never decoded. (Migrated 1:1 from `vulkan_video.rs`; the hand-replicated
+    /// sweep is now `plan_slot_recovery` itself.)
+    #[test]
+    fn taint_sweep_excludes_slots_from_an_earlier_loss() {
+        // Slots hold wires 0..7. Loss 1 starts at wire 4, so wires 4..7 are undecodable at the
+        // client. A second loss report arrives at wire 6 while they are all still resident.
+        let tainted_wires = [4i64, 5, 6, 7];
+
+        // WITHOUT the sweep this is the bug: the newest wire below 6 is wire 5 — squarely inside
+        // loss 1's unrepaired window — and it would be served as the "known-good" anchor.
+        let unswept = [0i64, 1, 2, 3, 4, 5, 6, 7];
+        let (_, picked_wire) = pick_anchor(&view(&unswept), 6).expect("unswept picks something");
+        assert!(
+            tainted_wires.contains(&picked_wire),
+            "precondition: without the sweep the anchor comes from the earlier loss window"
+        );
+
+        // WITH the plan, loss 1 taints 4..7 (and anchors on wire 3 — never a tainted wire, by
+        // the disjoint predicates), so loss 2 can only reach genuinely clean wires.
+        let mut wires = unswept;
+        let plan = plan_slot_recovery(&view(&wires), 4);
+        assert_eq!(plan.tainted, 0b1111_0000);
+        assert_eq!(plan.anchor, Some((3, 3)));
+        apply(&mut wires, plan.tainted);
+        assert_eq!(wires, [0, 1, 2, 3, -1, -1, -1, -1]);
+        let (slot, wire) = pick_anchor(&view(&wires), 6).expect("clean wires remain");
+        assert_eq!((slot, wire), (3, 3), "newest clean survivor is wire 3");
+
+        // Encoding resumes after recovery; wires 8..11 refill the swept slots and are clean. A
+        // later loss at wire 10 legitimately anchors on wire 9 — the sweep must not over-reject.
+        wires[4] = 8;
+        wires[5] = 9;
+        wires[6] = 10;
+        wires[7] = 11;
+        let plan = plan_slot_recovery(&view(&wires), 10);
+        assert_eq!(plan.anchor, Some((5, 9)), "wire 9 is post-recovery, clean");
+        apply(&mut wires, plan.tainted);
+
+        // A loss covering every live wire leaves nothing clean → decline, caller serves an IDR.
+        let mut all = [5i64, 6, 7, 8, 9, 10, 11, 12];
+        let plan = plan_slot_recovery(&view(&all), 5);
+        assert_eq!(plan.tainted, 0b1111_1111);
+        assert_eq!(plan.anchor, None);
+        apply(&mut all, plan.tainted);
+        assert_eq!(pick_anchor(&view(&all), 5), None);
+    }
+}
@@ -51,6 +51,24 @@ pub struct OpenH264Encoder {
 // whole value to that thread is therefore sound — there is no concurrent access to the handle.
 unsafe impl Send for OpenH264Encoder {}

+/// openh264's own ceiling: level 5.2, so 3840x2160 landscape or 2160x3840 portrait.
+///
+/// The long edge may reach 3840 and the short edge 2160 — the rule is orientation-aware, not a
+/// per-axis `w <= 3840 && h <= 2160`, so a portrait 2160x3840 session is legal.
+const OPENH264_MAX_LONG_EDGE: u32 = 3840;
+const OPENH264_MAX_SHORT_EDGE: u32 = 2160;
+
+/// Whether the bundled openh264 can encode this resolution at all.
+///
+/// Mirrors the check inside the crate we ship (openh264 0.9.3, `encoder.rs` `reinit`). That check
+/// runs on the FIRST ENCODE, not at encoder construction — so without this gate a too-large mode
+/// opens perfectly and then fails *every* submit, and the session connects and never delivers a
+/// frame. `Codec::max_dimension` does not cover it: it is keyed on the codec, and H.264 legitimately
+/// reaches 4096 on every hardware backend — this ceiling belongs to the software backend alone.
+fn openh264_supports_dimensions(width: u32, height: u32) -> bool {
+    width.max(height) <= OPENH264_MAX_LONG_EDGE && width.min(height) <= OPENH264_MAX_SHORT_EDGE
+}
+
 impl OpenH264Encoder {
    pub fn open(
        format: PixelFormat,
@@ -59,7 +77,16 @@ impl OpenH264Encoder {
        fps: u32,
        bitrate_bps: u64,
    ) -> Result<Self> {
-        // validate_dimensions() ran in open_video: even, non-zero, <= 4096.
+        // validate_dimensions() ran in open_video: even, non-zero, <= 4096. That leaves modes this
+        // encoder cannot serve (e.g. a legal 4096-wide H.264 mode), so refuse them here — at the
+        // open, where the caller still gets a real error — rather than at every submit.
+        ensure!(
+            openh264_supports_dimensions(width, height),
+            "openh264 cannot encode {width}x{height}: the software encoder tops out at \
+             {OPENH264_MAX_LONG_EDGE}x{OPENH264_MAX_SHORT_EDGE} (or \
+             {OPENH264_MAX_SHORT_EDGE}x{OPENH264_MAX_LONG_EDGE} portrait) — lower the client \
+             resolution, or use a host with a hardware encoder"
+        );
        let bps: u32 = bitrate_bps.try_into().unwrap_or(u32::MAX);
        let cfg = EncoderConfig::new()
            .usage_type(UsageType::ScreenContentRealTime)
@@ -336,4 +363,40 @@ mod tests {
            .any(|w| w[0] == 0 && w[1] == 0 && w[2] == 0 && w[3] == 1 && (w[4] & 0x1f) == 7);
        assert!(has_sps, "IDR must carry an SPS NAL (type 7)");
    }
+
+    /// The modes the software encoder can actually serve — including the portrait orientation,
+    /// which a naive per-axis `w <= 3840 && h <= 2160` would wrongly reject.
+    #[test]
+    fn openh264_accepts_up_to_4k_in_either_orientation() {
+        assert!(openh264_supports_dimensions(1920, 1080));
+        assert!(openh264_supports_dimensions(3840, 2160));
+        assert!(openh264_supports_dimensions(2160, 3840));
+        assert!(openh264_supports_dimensions(1080, 1920));
+    }
+
+    /// Modes `validate_dimensions` lets through (H.264 legitimately reaches 4096 on hardware) but
+    /// openh264 rejects on the first encode. Catching them at open is the whole point of the gate:
+    /// otherwise the session opens and then fails every single submit.
+    #[test]
+    fn openh264_rejects_modes_that_would_fail_on_first_submit() {
+        // 4096-wide is legal H.264 and passes `Codec::max_dimension`, but exceeds the long edge.
+        assert!(!openh264_supports_dimensions(4096, 2160));
+        assert!(!openh264_supports_dimensions(2160, 4096));
+        // Long edge is fine, short edge is not (e.g. an ultrawide-tall composite desktop).
+        assert!(!openh264_supports_dimensions(3840, 2400));
+        assert!(!openh264_supports_dimensions(2400, 3840));
+    }
+
+    /// A too-large mode must fail at `open`, not silently at every `submit`.
+    #[test]
+    fn open_refuses_a_mode_openh264_cannot_encode() {
+        // Matched rather than `expect_err`: `OpenH264Encoder` is not `Debug` (it wraps a raw C
+        // handle), so `expect_err` would not compile.
+        let err = match OpenH264Encoder::open(PixelFormat::Bgra, 4096, 2160, 60, 20_000_000) {
+            Ok(_) => panic!("4096x2160 exceeds openh264's long-edge ceiling and must be refused"),
+            Err(e) => e,
+        };
+        let msg = format!("{err:#}");
+        assert!(msg.contains("openh264 cannot encode 4096x2160"), "{msg}");
+    }
 }
@@ -66,425 +66,11 @@ use windows::Win32::Storage::FileSystem::{
 };
 use windows::Win32::System::LibraryLoader::GetModuleFileNameW;

-// ---------------------------------------------------------------------------------------------
-// Mirrored AMF C ABI (written against GPUOpen header release v1.4.36 — amf/public/include; every
-// slot below is a base-interface slot whose layout is stable since <= v1.4.34, the loader's
-// accepted ABI floor, so the mirror is valid on every runtime the loader admits).
-//
-// Layout rules this mirror relies on: every AMF interface is a struct whose sole member is a
-// pointer to a C vtable; derived interfaces PREPEND their base's slots in order (AMFInterface →
-// AMFPropertyStorage → AMFData → AMFBuffer/AMFSurface), so a derived pointer is usable through a
-// base vtable mirror. Slots we never call are declared as bare `*const c_void` placeholders —
-// same size/alignment as the function pointer they stand in for. `AMF_STD_CALL` is `__stdcall`
-// (Rust `extern "system"`); the two DLL entry points are `__cdecl` (`extern "C"`); on x86_64
-// both collapse to the one Windows calling convention.
-// ---------------------------------------------------------------------------------------------
-mod sys {
-    use std::ffi::c_void;
-
-    /// `AMF_RESULT` (core/Result.h) — a plain C enum, sequential from 0. Only the codes this
-    /// module branches on are named; everything else is reported numerically via [`result_name`].
-    pub type AmfResult = i32;
-    pub const AMF_OK: AmfResult = 0;
-    pub const AMF_EOF: AmfResult = 23;
-    pub const AMF_REPEAT: AmfResult = 24;
-    pub const AMF_INPUT_FULL: AmfResult = 25;
-    pub const AMF_NEED_MORE_INPUT: AmfResult = 44;
-
-    /// Human-readable name for an `AMF_RESULT` (diagnostics only — the numeric value rides along
-    /// so an unnamed code is still identifiable against Result.h).
-    pub fn result_name(r: AmfResult) -> &'static str {
-        match r {
-            0 => "AMF_OK",
-            1 => "AMF_FAIL",
-            2 => "AMF_UNEXPECTED",
-            3 => "AMF_ACCESS_DENIED",
-            4 => "AMF_INVALID_ARG",
-            5 => "AMF_OUT_OF_RANGE",
-            6 => "AMF_OUT_OF_MEMORY",
-            7 => "AMF_INVALID_POINTER",
-            8 => "AMF_NO_INTERFACE",
-            9 => "AMF_NOT_IMPLEMENTED",
-            10 => "AMF_NOT_SUPPORTED",
-            11 => "AMF_NOT_FOUND",
-            12 => "AMF_ALREADY_INITIALIZED",
-            13 => "AMF_NOT_INITIALIZED",
-            14 => "AMF_INVALID_FORMAT",
-            15 => "AMF_WRONG_STATE",
-            17 => "AMF_NO_DEVICE",
-            18 => "AMF_DIRECTX_FAILED",
-            23 => "AMF_EOF",
-            24 => "AMF_REPEAT",
-            25 => "AMF_INPUT_FULL",
-            26 => "AMF_RESOLUTION_CHANGED",
-            28 => "AMF_INVALID_DATA_TYPE",
-            29 => "AMF_INVALID_RESOLUTION",
-            30 => "AMF_CODEC_NOT_SUPPORTED",
-            31 => "AMF_SURFACE_FORMAT_NOT_SUPPORTED",
-            32 => "AMF_SURFACE_MUST_BE_SHARED",
-            36 => "AMF_ENCODER_NOT_PRESENT",
-            44 => "AMF_NEED_MORE_INPUT",
-            _ => "AMF_<unnamed>",
-        }
-    }
-
-    /// The AMF header release this FFI mirror was written against: `AMF_FULL_VERSION` for 1.4.36.0
-    /// (core/Version.h `AMF_MAKE_FULL_VERSION`). This is the version claimed to `AMFInit` — but
-    /// capped at the runtime's own reported version (see `load_factory`), so an older-but-accepted
-    /// runtime is asked only for the ABI it actually provides.
-    pub const AMF_HEADER_VERSION: u64 = (1u64 << 48) | (4u64 << 32) | (36u64 << 16);
-
-    /// The oldest AMF runtime the loader accepts (`AMF_FULL_VERSION` 1.4.34.0 — AMD Adrenalin
-    /// 24.6.1). This is an **ABI floor, not a feature floor**: every vtable slot mirrored in this
-    /// module belongs to a base interface (`AMFFactory`/`AMFContext`/`AMFComponent`/`AMFData`/
-    /// `AMFBuffer`) whose layout has been stable — append-only, no mid-vtable insertions — since
-    /// well before 1.4.34, so a 1.4.34 runtime is guaranteed to expose every mirrored slot at its
-    /// mirrored offset. Everything 1.4.35/1.4.36 added that this path can touch (new HQ presets,
-    /// AV1 B-frame / picture management) is a *string-keyed encoder property*, applied through
-    /// [`set_prop`] with `required=false` — a runtime that lacks it rejects the property (logged)
-    /// and the feature degrades, rather than shifting any vtable offset. Below this floor the
-    /// mirror is not guaranteed to match, so the loader declines cleanly (an old-driver decline,
-    /// never UB).
-    pub const AMF_MIN_VERSION: u64 = (1u64 << 48) | (4u64 << 32) | (34u64 << 16);
-
-    /// `AMF_SURFACE_FORMAT` (core/Surface.h).
-    pub const AMF_SURFACE_NV12: i32 = 1;
-    pub const AMF_SURFACE_P010: i32 = 10;
-
-    /// `AMF_DX_VERSION::AMF_DX11_1` (core/Data.h) — the `InitDX11` version argument.
-    pub const AMF_DX11_1: i32 = 111;
-
-    /// `AMF_MEMORY_TYPE::AMF_MEMORY_HOST` (core/Data.h) — the `AllocBuffer` memory type for the
-    /// CPU-filled HDR-metadata buffer.
-    pub const AMF_MEMORY_HOST: i32 = 1;
-
-    /// `AMFHDRMetadata` (components/ColorSpace.h) — the payload of the `*InHDRMetadata` encoder
-    /// property (an `AMFBuffer` holding exactly this struct). Same units as the HEVC ST.2086 SEI
-    /// and [`punktfunk_core::quic::HdrMeta`]: chromaticities in 1/50000, mastering luminance in
-    /// 0.0001 cd/m², CLL/FALL in nits. 28 bytes, no padding.
-    #[repr(C)]
-    pub struct AmfHdrMetadata {
-        pub red_primary: [u16; 2],
-        pub green_primary: [u16; 2],
-        pub blue_primary: [u16; 2],
-        pub white_point: [u16; 2],
-        pub max_mastering_luminance: u32,
-        pub min_mastering_luminance: u32,
-        pub max_content_light_level: u16,
-        pub max_frame_average_light_level: u16,
-    }
-
-    /// `AMFGuid` (core/Platform.h) — data41..data48 flattened into an array (identical layout).
-    #[repr(C)]
-    pub struct AmfGuid {
-        pub data1: u32,
-        pub data2: u16,
-        pub data3: u16,
-        pub data4: [u8; 8],
-    }
-
-    /// `IID_AMFBuffer` (core/Buffer.h `AMF_DECLARE_IID`) — for `QueryInterface` on the encoder's
-    /// output `AMFData` to reach `GetNative`/`GetSize`.
-    pub const IID_AMF_BUFFER: AmfGuid = AmfGuid {
-        data1: 0xb04b_7248,
-        data2: 0xb6f0,
-        data3: 0x4321,
-        data4: [0xb6, 0x91, 0xba, 0xa4, 0x74, 0x0f, 0x9f, 0xcb],
-    };
-
-    // `AMF_VARIANT_TYPE` (core/Variant.h) — the tags this module writes/reads.
-    pub const AMF_VARIANT_BOOL: i32 = 1;
-    pub const AMF_VARIANT_INT64: i32 = 2;
-    pub const AMF_VARIANT_RATE: i32 = 7;
-    pub const AMF_VARIANT_INTERFACE: i32 = 12;
-
-    /// `AMFVariantStruct` (core/Variant.h): a 4-byte C-enum tag + a 16-byte union whose largest
-    /// members are pointer/`amf_int64`/`AMFFloatVector4D` (align 8) → 24 bytes total, tag at 0,
-    /// payload at 8. Passed BY VALUE to `SetProperty` (Win64 passes >8-byte aggregates by hidden
-    /// reference on both sides, so declaring it by value matches the C compiler). The payload is
-    /// stored as two fully-initialised `u64`s — little-endian packing puts a bool in byte 0, an
-    /// `amf_int64` in word 0, and an `AMFRate{num,den}` as `num | den << 32`, exactly the union's
-    /// in-memory layout — so no partially-initialised union bytes ever cross the FFI.
-    #[repr(C)]
-    pub struct AmfVariant {
-        pub vtype: i32,
-        pub payload: [u64; 2],
-    }
-
-    impl AmfVariant {
-        pub fn zeroed() -> Self {
-            AmfVariant {
-                vtype: 0, // AMF_VARIANT_EMPTY
-                payload: [0, 0],
-            }
-        }
-        pub fn from_i64(v: i64) -> Self {
-            AmfVariant {
-                vtype: AMF_VARIANT_INT64,
-                payload: [v as u64, 0],
-            }
-        }
-        pub fn from_bool(v: bool) -> Self {
-            AmfVariant {
-                vtype: AMF_VARIANT_BOOL,
-                payload: [v as u64, 0],
-            }
-        }
-        /// `AMFRate { num, den }` — two little-endian `amf_uint32`s in the union's first 8 bytes.
-        pub fn from_rate(num: u32, den: u32) -> Self {
-            AmfVariant {
-                vtype: AMF_VARIANT_RATE,
-                payload: [num as u64 | ((den as u64) << 32), 0],
-            }
-        }
-        /// An `AMFInterface*` payload (`pInterface` in the union's first 8 bytes). The property
-        /// storage AddRefs the interface when it copies the variant in (the C++ template
-        /// `SetProperty(name, AMFVariant(value))` passes a temporary whose destructor releases,
-        /// so `SetProperty` must take its own reference) — the caller keeps sole ownership of the
-        /// reference it already holds.
-        pub fn from_interface(p: *mut c_void) -> Self {
-            AmfVariant {
-                vtype: AMF_VARIANT_INTERFACE,
-                payload: [p as usize as u64, 0],
-            }
-        }
-        /// Read back an `amf_int64` payload (only valid when `vtype == AMF_VARIANT_INT64`).
-        pub fn as_i64(&self) -> Option<i64> {
-            (self.vtype == AMF_VARIANT_INT64).then_some(self.payload[0] as i64)
-        }
-    }
-
-    /// Placeholder for a vtable slot this module never calls — same size/align as the function
-    /// pointer it stands in for, present only to keep the following slots at their C offsets.
-    pub type Slot = *const c_void;
-
-    // -- AMFFactory (core/Factory.h; NOT refcounted — a process singleton) ----------------------
-    #[repr(C)]
-    pub struct AmfFactory {
-        pub vtbl: *const AmfFactoryVtbl,
-    }
-    #[repr(C)]
-    pub struct AmfFactoryVtbl {
-        pub create_context:
-            unsafe extern "system" fn(*mut AmfFactory, *mut *mut AmfContext) -> AmfResult,
-        pub create_component: unsafe extern "system" fn(
-            *mut AmfFactory,
-            *mut AmfContext,
-            *const u16,
-            *mut *mut AmfComponent,
-        ) -> AmfResult,
-        pub set_cache_folder: Slot,
-        pub get_cache_folder: Slot,
-        pub get_debug: Slot,
-        pub get_trace: Slot,
-        pub get_programs: Slot,
-    }
-
-    // -- AMFContext (core/Context.h) ------------------------------------------------------------
-    #[repr(C)]
-    pub struct AmfContext {
-        pub vtbl: *const AmfContextVtbl,
-    }
-    #[repr(C)]
-    pub struct AmfContextVtbl {
-        // AMFInterface
-        pub acquire: Slot,
-        pub release: unsafe extern "system" fn(*mut AmfContext) -> i32,
-        pub query_interface: Slot,
-        // AMFPropertyStorage
-        pub set_property: Slot,
-        pub get_property: Slot,
-        pub has_property: Slot,
-        pub get_property_count: Slot,
-        pub get_property_at: Slot,
-        pub clear: Slot,
-        pub add_to: Slot,
-        pub copy_to: Slot,
-        pub add_observer: Slot,
-        pub remove_observer: Slot,
-        // AMFContext
-        pub terminate: unsafe extern "system" fn(*mut AmfContext) -> AmfResult,
-        pub init_dx9: Slot,
-        pub get_dx9_device: Slot,
-        pub lock_dx9: Slot,
-        pub unlock_dx9: Slot,
-        pub init_dx11: unsafe extern "system" fn(*mut AmfContext, *mut c_void, i32) -> AmfResult,
-        pub get_dx11_device: Slot,
-        pub lock_dx11: Slot,
-        pub unlock_dx11: Slot,
-        pub init_opencl: Slot,
-        pub get_opencl_context: Slot,
-        pub get_opencl_command_queue: Slot,
-        pub get_opencl_device_id: Slot,
-        pub get_opencl_compute_factory: Slot,
-        pub init_opencl_ex: Slot,
-        pub lock_opencl: Slot,
-        pub unlock_opencl: Slot,
-        pub init_opengl: Slot,
-        pub get_opengl_context: Slot,
-        pub get_opengl_drawable: Slot,
-        pub lock_opengl: Slot,
-        pub unlock_opengl: Slot,
-        pub init_xv: Slot,
-        pub get_xv_device: Slot,
-        pub lock_xv: Slot,
-        pub unlock_xv: Slot,
-        pub init_gralloc: Slot,
-        pub get_gralloc_device: Slot,
-        pub lock_gralloc: Slot,
-        pub unlock_gralloc: Slot,
-        pub alloc_buffer: unsafe extern "system" fn(
-            *mut AmfContext,
-            i32, // AMF_MEMORY_TYPE
-            usize,
-            *mut *mut AmfBuffer,
-        ) -> AmfResult,
-        pub alloc_surface: Slot,
-        pub alloc_audio_buffer: Slot,
-        pub create_buffer_from_host_native: Slot,
-        pub create_surface_from_host_native: Slot,
-        pub create_surface_from_dx9_native: Slot,
-        /// Out-param is `AMFSurface**` in the header; declared as the `AmfData` base here because
-        /// every surface call this module makes (`SetPts`, `SetProperty`, `Release`,
-        /// `SubmitInput`) lives in the `AMFData` vtable prefix, which `AMFSurfaceVtbl` reproduces
-        /// slot-for-slot (single inheritance, same object pointer).
-        pub create_surface_from_dx11_native: unsafe extern "system" fn(
-            *mut AmfContext,
-            *mut c_void,
-            *mut *mut AmfData,
-            *mut c_void,
-        ) -> AmfResult,
-        pub create_surface_from_opengl_native: Slot,
-        pub create_surface_from_gralloc_native: Slot,
-        pub create_surface_from_opencl_native: Slot,
-        pub create_buffer_from_opencl_native: Slot,
-        pub get_compute: Slot,
-    }
-
-    // -- AMFComponent (components/Component.h) --------------------------------------------------
-    #[repr(C)]
-    pub struct AmfComponent {
-        pub vtbl: *const AmfComponentVtbl,
-    }
-    #[repr(C)]
-    pub struct AmfComponentVtbl {
-        // AMFInterface
-        pub acquire: Slot,
-        pub release: unsafe extern "system" fn(*mut AmfComponent) -> i32,
-        pub query_interface: Slot,
-        // AMFPropertyStorage
-        pub set_property:
-            unsafe extern "system" fn(*mut AmfComponent, *const u16, AmfVariant) -> AmfResult,
-        pub get_property: Slot,
-        pub has_property: Slot,
-        pub get_property_count: Slot,
-        pub get_property_at: Slot,
-        pub clear: Slot,
-        pub add_to: Slot,
-        pub copy_to: Slot,
-        pub add_observer: Slot,
-        pub remove_observer: Slot,
-        // AMFPropertyStorageEx
-        pub get_properties_info_count: Slot,
-        pub get_property_info_at: Slot,
-        pub get_property_info: Slot,
-        pub validate_property: Slot,
-        // AMFComponent
-        pub init: unsafe extern "system" fn(*mut AmfComponent, i32, i32, i32) -> AmfResult,
-        pub reinit: Slot,
-        pub terminate: unsafe extern "system" fn(*mut AmfComponent) -> AmfResult,
-        pub drain: unsafe extern "system" fn(*mut AmfComponent) -> AmfResult,
-        pub flush: unsafe extern "system" fn(*mut AmfComponent) -> AmfResult,
-        pub submit_input: unsafe extern "system" fn(*mut AmfComponent, *mut AmfData) -> AmfResult,
-        pub query_output:
-            unsafe extern "system" fn(*mut AmfComponent, *mut *mut AmfData) -> AmfResult,
-        pub get_context: Slot,
-        pub set_output_data_allocator_cb: Slot,
-        pub get_caps: Slot,
-        pub optimize: Slot,
-    }
-
-    // -- AMFData (core/Data.h) — also the usable prefix of AMFSurface --------------------------
-    #[repr(C)]
-    pub struct AmfData {
-        pub vtbl: *const AmfDataVtbl,
-    }
-    #[repr(C)]
-    pub struct AmfDataVtbl {
-        // AMFInterface
-        pub acquire: Slot,
-        pub release: unsafe extern "system" fn(*mut AmfData) -> i32,
-        pub query_interface:
-            unsafe extern "system" fn(*mut AmfData, *const AmfGuid, *mut *mut c_void) -> AmfResult,
-        // AMFPropertyStorage
-        pub set_property:
-            unsafe extern "system" fn(*mut AmfData, *const u16, AmfVariant) -> AmfResult,
-        pub get_property:
-            unsafe extern "system" fn(*mut AmfData, *const u16, *mut AmfVariant) -> AmfResult,
-        pub has_property: Slot,
-        pub get_property_count: Slot,
-        pub get_property_at: Slot,
-        pub clear: Slot,
-        pub add_to: Slot,
-        pub copy_to: Slot,
-        pub add_observer: Slot,
-        pub remove_observer: Slot,
-        // AMFData
-        pub get_memory_type: Slot,
-        pub duplicate: Slot,
-        pub convert: Slot,
-        pub interop: Slot,
-        pub get_data_type: Slot,
-        pub is_reusable: Slot,
-        pub set_pts: unsafe extern "system" fn(*mut AmfData, i64),
-        pub get_pts: Slot,
-        pub set_duration: Slot,
-        pub get_duration: Slot,
-    }
-
-    // -- AMFBuffer (core/Buffer.h) — the encoder's output object -------------------------------
-    #[repr(C)]
-    pub struct AmfBuffer {
-        pub vtbl: *const AmfBufferVtbl,
-    }
-    #[repr(C)]
-    pub struct AmfBufferVtbl {
-        // AMFInterface + AMFPropertyStorage + AMFData prefix (identical order to AmfDataVtbl).
-        pub acquire: Slot,
-        pub release: unsafe extern "system" fn(*mut AmfBuffer) -> i32,
-        pub query_interface: Slot,
-        pub set_property: Slot,
-        pub get_property: Slot,
-        pub has_property: Slot,
-        pub get_property_count: Slot,
-        pub get_property_at: Slot,
-        pub clear: Slot,
-        pub add_to: Slot,
-        pub copy_to: Slot,
-        pub add_observer: Slot,
-        pub remove_observer: Slot,
-        pub get_memory_type: Slot,
-        pub duplicate: Slot,
-        pub convert: Slot,
-        pub interop: Slot,
-        pub get_data_type: Slot,
-        pub is_reusable: Slot,
-        pub set_pts: Slot,
-        pub get_pts: Slot,
-        pub set_duration: Slot,
-        pub get_duration: Slot,
-        // AMFBuffer
-        pub set_size: Slot,
-        pub get_size: unsafe extern "system" fn(*mut AmfBuffer) -> usize,
-        pub get_native: unsafe extern "system" fn(*mut AmfBuffer) -> *mut c_void,
-        pub add_observer_buffer: Slot,
-        pub remove_observer_buffer: Slot,
-    }
-
-    // -- DLL entry points (core/Factory.h; AMF_CDECL_CALL) --------------------------------------
-    pub type AmfQueryVersionFn = unsafe extern "C" fn(*mut u64) -> AmfResult;
-    pub type AmfInitFn = unsafe extern "C" fn(u64, *mut *mut AmfFactory) -> AmfResult;
-}
+// The hand-mirrored AMF C ABI lives in `amf_sys.rs` (WP7.4): pure FFI surface, no policy —
+// isolating the unsafe vtable mirror from the encoder logic is the crate's stated review
+// goal, and the `#[path]` keeps the module name (`sys::`) and every call site unchanged.
+#[path = "amf_sys.rs"]
+mod sys;

 use sys::{result_name, AmfVariant};

@@ -1272,7 +858,11 @@ impl AmfEncoder {
        if codec == Codec::Av1 && !probe_can_encode(Codec::Av1) {
            bail!("this GPU/driver declined AV1 encode (RDNA3+ required) — native AMF probe");
        }
-        let ten_bit = bit_depth >= 10 || matches!(format, PixelFormat::P010 | PixelFormat::Rgb10a2);
+        // Depth follows the delivered pixels, not the negotiated depth ([`crate::ten_bit_input`]).
+        // With the old `bit_depth >= 10 || …` shape a 10-bit-negotiated session over an 8-bit
+        // capture derived `expected = P010`, tripped the format check below and ended the session
+        // at open — on exactly the configuration the capturer had already downgraded on purpose.
+        let ten_bit = crate::ten_bit_input(format, bit_depth);
        // Zero-copy by construction: the input ring is NV12/P010 fed by same-format
        // CopySubresourceRegion. Any other capture format (Bgra/Rgb10a2 video-processor fallback,
        // CPU frames) has no native input path — and since Phase 3 no ffmpeg readback to degrade
@@ -2369,31 +1959,27 @@ impl Encoder for AmfEncoder {
        if !self.ltr_active || first < 0 || first > last {
            return false;
        }
-        // Taint sweep BEFORE picking the anchor: an LTR marked at-or-after the loss start was
-        // encoded inside the client's corrupt window — the client either never received it or
-        // decoded it against a broken reference chain. Serving it as "known-good" on a LATER
-        // loss ships corruption as the recovery anchor (and every subsequent mark re-samples
-        // it). Dropped slots stay dropped; the cadence re-marks a clean frame within ~1/4 s.
-        for marked in self.ltr_slots.iter_mut() {
-            if marked.is_some_and(|idx| idx >= first) {
+        // The taint-sweep + anchor-pick POLICY lives in `rfi::plan_slot_recovery` (one decision
+        // shared with QSV and Vulkan Video); this backend's mechanism is: distrust = clear the
+        // mirror slot (dropped slots stay dropped; the cadence re-marks a clean frame within
+        // ~1/4 s). `ltr_slots` store the WIRE frame index of the marked frame (`submit_indexed`
+        // pins `frame_idx` to it per submission), so they compare directly against the client's
+        // `first` — and stay comparable across encoder rebuilds/resets, where an internal counter
+        // would make the pre-loss check vacuous and risk force-referencing an LTR marked INSIDE
+        // the lost range.
+        let view: Vec<(usize, i64)> = self
+            .ltr_slots
+            .iter()
+            .enumerate()
+            .filter_map(|(s, m)| m.map(|w| (s, w)))
+            .collect();
+        let plan = super::rfi::plan_slot_recovery(&view, first);
+        for (slot, marked) in self.ltr_slots.iter_mut().enumerate() {
+            if plan.tainted & (1 << slot) != 0 {
                *marked = None;
            }
        }
-        // Pick the newest LTR strictly OLDER than the loss: the most recent known-good reference the
-        // client still holds, so re-referencing it costs the least (smallest recovery-frame residual).
-        // `ltr_slots` store the WIRE frame index of the marked frame (`submit_indexed` pins
-        // `frame_idx` to it per submission), so they compare directly against the client's `first`
-        // — and stay comparable across encoder rebuilds/resets, where an internal counter would
-        // make this check vacuous and risk force-referencing an LTR marked INSIDE the lost range.
-        let mut best: Option<(usize, i64)> = None;
-        for (slot, marked) in self.ltr_slots.iter().enumerate() {
-            if let Some(idx) = *marked {
-                if idx < first && best.is_none_or(|(_, b)| idx > b) {
-                    best = Some((slot, idx));
-                }
-            }
-        }
-        match best {
+        match plan.anchor {
            Some((slot, ltr_frame)) => {
                // Queue the force for the next submit; that frame ships tagged `recovery_anchor`.
                self.pending_force = Some(slot);
@@ -2422,6 +2008,8 @@ impl Encoder for AmfEncoder {

    fn caps(&self) -> EncoderCaps {
        EncoderCaps {
+            // As Windows NVENC: the capturer composites; this backend never reads `frame.cursor`.
+            blends_cursor: false,
            // LTR-RFI: AMD's reference invalidation is the user long-term-reference path (mark a
            // frame, force a later one to re-reference it). True only when the live driver accepted
            // the LTR slots at open — otherwise loss recovery falls back to a full IDR.
@@ -2760,6 +2348,10 @@ mod tests {
    }

    /// The `p`-quantile of `samples` (µs), sorting in place. `0` when empty.
+    /// Gated like its only caller, the `amf-qsv`-only §5.2 latency A/B below — otherwise a
+    /// `--features nvenc,qsv` build compiles this helper with the benchmark cfg'd out and trips
+    /// `dead_code` (which the crate root no longer blanket-allows).
+    #[cfg(feature = "amf-qsv")]
    fn percentile(samples: &mut [u128], p: f64) -> u128 {
        if samples.is_empty() {
            return 0;
@@ -2778,6 +2370,8 @@ mod tests {
    /// so its submit→AU is the bare ASIC time. The last ~2 unflushed frames on the ffmpeg path
    /// are left unmeasured (dropped with the encoder) so every recorded sample is a genuine paced
    /// submit→AU.
+    /// Gated like its only caller (see [`percentile`]).
+    #[cfg(feature = "amf-qsv")]
    #[allow(clippy::too_many_arguments)]
    fn drive_and_measure(
        enc: &mut dyn Encoder,
@@ -0,0 +1,414 @@
+//! Mirrored AMF C ABI (written against GPUOpen header release v1.4.36 — amf/public/include; every
+//! slot below is a base-interface slot whose layout is stable since <= v1.4.34, the loader's
+//! accepted ABI floor, so the mirror is valid on every runtime the loader admits).
+//!
+//! Layout rules this mirror relies on: every AMF interface is a struct whose sole member is a
+//! pointer to a C vtable; derived interfaces PREPEND their base's slots in order (AMFInterface →
+//! AMFPropertyStorage → AMFData → AMFBuffer/AMFSurface), so a derived pointer is usable through a
+//! base vtable mirror. Slots we never call are declared as bare `*const c_void` placeholders —
+//! same size/alignment as the function pointer they stand in for. `AMF_STD_CALL` is `__stdcall`
+//! (Rust `extern "system"`); the two DLL entry points are `__cdecl` (`extern "C"`); on x86_64
+//! both collapse to the one Windows calling convention.
+
+use std::ffi::c_void;
+
+/// `AMF_RESULT` (core/Result.h) — a plain C enum, sequential from 0. Only the codes this
+/// module branches on are named; everything else is reported numerically via [`result_name`].
+pub type AmfResult = i32;
+pub const AMF_OK: AmfResult = 0;
+pub const AMF_EOF: AmfResult = 23;
+pub const AMF_REPEAT: AmfResult = 24;
+pub const AMF_INPUT_FULL: AmfResult = 25;
+pub const AMF_NEED_MORE_INPUT: AmfResult = 44;
+
+/// Human-readable name for an `AMF_RESULT` (diagnostics only — the numeric value rides along
+/// so an unnamed code is still identifiable against Result.h).
+pub fn result_name(r: AmfResult) -> &'static str {
+    match r {
+        0 => "AMF_OK",
+        1 => "AMF_FAIL",
+        2 => "AMF_UNEXPECTED",
+        3 => "AMF_ACCESS_DENIED",
+        4 => "AMF_INVALID_ARG",
+        5 => "AMF_OUT_OF_RANGE",
+        6 => "AMF_OUT_OF_MEMORY",
+        7 => "AMF_INVALID_POINTER",
+        8 => "AMF_NO_INTERFACE",
+        9 => "AMF_NOT_IMPLEMENTED",
+        10 => "AMF_NOT_SUPPORTED",
+        11 => "AMF_NOT_FOUND",
+        12 => "AMF_ALREADY_INITIALIZED",
+        13 => "AMF_NOT_INITIALIZED",
+        14 => "AMF_INVALID_FORMAT",
+        15 => "AMF_WRONG_STATE",
+        17 => "AMF_NO_DEVICE",
+        18 => "AMF_DIRECTX_FAILED",
+        23 => "AMF_EOF",
+        24 => "AMF_REPEAT",
+        25 => "AMF_INPUT_FULL",
+        26 => "AMF_RESOLUTION_CHANGED",
+        28 => "AMF_INVALID_DATA_TYPE",
+        29 => "AMF_INVALID_RESOLUTION",
+        30 => "AMF_CODEC_NOT_SUPPORTED",
+        31 => "AMF_SURFACE_FORMAT_NOT_SUPPORTED",
+        32 => "AMF_SURFACE_MUST_BE_SHARED",
+        36 => "AMF_ENCODER_NOT_PRESENT",
+        44 => "AMF_NEED_MORE_INPUT",
+        _ => "AMF_<unnamed>",
+    }
+}
+
+/// The AMF header release this FFI mirror was written against: `AMF_FULL_VERSION` for 1.4.36.0
+/// (core/Version.h `AMF_MAKE_FULL_VERSION`). This is the version claimed to `AMFInit` — but
+/// capped at the runtime's own reported version (see `load_factory`), so an older-but-accepted
+/// runtime is asked only for the ABI it actually provides.
+pub const AMF_HEADER_VERSION: u64 = (1u64 << 48) | (4u64 << 32) | (36u64 << 16);
+
+/// The oldest AMF runtime the loader accepts (`AMF_FULL_VERSION` 1.4.34.0 — AMD Adrenalin
+/// 24.6.1). This is an **ABI floor, not a feature floor**: every vtable slot mirrored in this
+/// module belongs to a base interface (`AMFFactory`/`AMFContext`/`AMFComponent`/`AMFData`/
+/// `AMFBuffer`) whose layout has been stable — append-only, no mid-vtable insertions — since
+/// well before 1.4.34, so a 1.4.34 runtime is guaranteed to expose every mirrored slot at its
+/// mirrored offset. Everything 1.4.35/1.4.36 added that this path can touch (new HQ presets,
+/// AV1 B-frame / picture management) is a *string-keyed encoder property*, applied through
+/// [`set_prop`] with `required=false` — a runtime that lacks it rejects the property (logged)
+/// and the feature degrades, rather than shifting any vtable offset. Below this floor the
+/// mirror is not guaranteed to match, so the loader declines cleanly (an old-driver decline,
+/// never UB).
+pub const AMF_MIN_VERSION: u64 = (1u64 << 48) | (4u64 << 32) | (34u64 << 16);
+
+/// `AMF_SURFACE_FORMAT` (core/Surface.h).
+pub const AMF_SURFACE_NV12: i32 = 1;
+pub const AMF_SURFACE_P010: i32 = 10;
+
+/// `AMF_DX_VERSION::AMF_DX11_1` (core/Data.h) — the `InitDX11` version argument.
+pub const AMF_DX11_1: i32 = 111;
+
+/// `AMF_MEMORY_TYPE::AMF_MEMORY_HOST` (core/Data.h) — the `AllocBuffer` memory type for the
+/// CPU-filled HDR-metadata buffer.
+pub const AMF_MEMORY_HOST: i32 = 1;
+
+/// `AMFHDRMetadata` (components/ColorSpace.h) — the payload of the `*InHDRMetadata` encoder
+/// property (an `AMFBuffer` holding exactly this struct). Same units as the HEVC ST.2086 SEI
+/// and [`punktfunk_core::quic::HdrMeta`]: chromaticities in 1/50000, mastering luminance in
+/// 0.0001 cd/m², CLL/FALL in nits. 28 bytes, no padding.
+#[repr(C)]
+pub struct AmfHdrMetadata {
+    pub red_primary: [u16; 2],
+    pub green_primary: [u16; 2],
+    pub blue_primary: [u16; 2],
+    pub white_point: [u16; 2],
+    pub max_mastering_luminance: u32,
+    pub min_mastering_luminance: u32,
+    pub max_content_light_level: u16,
+    pub max_frame_average_light_level: u16,
+}
+
+/// `AMFGuid` (core/Platform.h) — data41..data48 flattened into an array (identical layout).
+#[repr(C)]
+pub struct AmfGuid {
+    pub data1: u32,
+    pub data2: u16,
+    pub data3: u16,
+    pub data4: [u8; 8],
+}
+
+/// `IID_AMFBuffer` (core/Buffer.h `AMF_DECLARE_IID`) — for `QueryInterface` on the encoder's
+/// output `AMFData` to reach `GetNative`/`GetSize`.
+pub const IID_AMF_BUFFER: AmfGuid = AmfGuid {
+    data1: 0xb04b_7248,
+    data2: 0xb6f0,
+    data3: 0x4321,
+    data4: [0xb6, 0x91, 0xba, 0xa4, 0x74, 0x0f, 0x9f, 0xcb],
+};
+
+// `AMF_VARIANT_TYPE` (core/Variant.h) — the tags this module writes/reads.
+pub const AMF_VARIANT_BOOL: i32 = 1;
+pub const AMF_VARIANT_INT64: i32 = 2;
+pub const AMF_VARIANT_RATE: i32 = 7;
+pub const AMF_VARIANT_INTERFACE: i32 = 12;
+
+/// `AMFVariantStruct` (core/Variant.h): a 4-byte C-enum tag + a 16-byte union whose largest
+/// members are pointer/`amf_int64`/`AMFFloatVector4D` (align 8) → 24 bytes total, tag at 0,
+/// payload at 8. Passed BY VALUE to `SetProperty` (Win64 passes >8-byte aggregates by hidden
+/// reference on both sides, so declaring it by value matches the C compiler). The payload is
+/// stored as two fully-initialised `u64`s — little-endian packing puts a bool in byte 0, an
+/// `amf_int64` in word 0, and an `AMFRate{num,den}` as `num | den << 32`, exactly the union's
+/// in-memory layout — so no partially-initialised union bytes ever cross the FFI.
+#[repr(C)]
+pub struct AmfVariant {
+    pub vtype: i32,
+    pub payload: [u64; 2],
+}
+
+impl AmfVariant {
+    pub fn zeroed() -> Self {
+        AmfVariant {
+            vtype: 0, // AMF_VARIANT_EMPTY
+            payload: [0, 0],
+        }
+    }
+    pub fn from_i64(v: i64) -> Self {
+        AmfVariant {
+            vtype: AMF_VARIANT_INT64,
+            payload: [v as u64, 0],
+        }
+    }
+    pub fn from_bool(v: bool) -> Self {
+        AmfVariant {
+            vtype: AMF_VARIANT_BOOL,
+            payload: [v as u64, 0],
+        }
+    }
+    /// `AMFRate { num, den }` — two little-endian `amf_uint32`s in the union's first 8 bytes.
+    pub fn from_rate(num: u32, den: u32) -> Self {
+        AmfVariant {
+            vtype: AMF_VARIANT_RATE,
+            payload: [num as u64 | ((den as u64) << 32), 0],
+        }
+    }
+    /// An `AMFInterface*` payload (`pInterface` in the union's first 8 bytes). The property
+    /// storage AddRefs the interface when it copies the variant in (the C++ template
+    /// `SetProperty(name, AMFVariant(value))` passes a temporary whose destructor releases,
+    /// so `SetProperty` must take its own reference) — the caller keeps sole ownership of the
+    /// reference it already holds.
+    pub fn from_interface(p: *mut c_void) -> Self {
+        AmfVariant {
+            vtype: AMF_VARIANT_INTERFACE,
+            payload: [p as usize as u64, 0],
+        }
+    }
+    /// Read back an `amf_int64` payload (only valid when `vtype == AMF_VARIANT_INT64`).
+    pub fn as_i64(&self) -> Option<i64> {
+        (self.vtype == AMF_VARIANT_INT64).then_some(self.payload[0] as i64)
+    }
+}
+
+/// Placeholder for a vtable slot this module never calls — same size/align as the function
+/// pointer it stands in for, present only to keep the following slots at their C offsets.
+pub type Slot = *const c_void;
+
+// -- AMFFactory (core/Factory.h; NOT refcounted — a process singleton) ----------------------
+#[repr(C)]
+pub struct AmfFactory {
+    pub vtbl: *const AmfFactoryVtbl,
+}
+#[repr(C)]
+pub struct AmfFactoryVtbl {
+    pub create_context:
+        unsafe extern "system" fn(*mut AmfFactory, *mut *mut AmfContext) -> AmfResult,
+    pub create_component: unsafe extern "system" fn(
+        *mut AmfFactory,
+        *mut AmfContext,
+        *const u16,
+        *mut *mut AmfComponent,
+    ) -> AmfResult,
+    pub set_cache_folder: Slot,
+    pub get_cache_folder: Slot,
+    pub get_debug: Slot,
+    pub get_trace: Slot,
+    pub get_programs: Slot,
+}
+
+// -- AMFContext (core/Context.h) ------------------------------------------------------------
+#[repr(C)]
+pub struct AmfContext {
+    pub vtbl: *const AmfContextVtbl,
+}
+#[repr(C)]
+pub struct AmfContextVtbl {
+    // AMFInterface
+    pub acquire: Slot,
+    pub release: unsafe extern "system" fn(*mut AmfContext) -> i32,
+    pub query_interface: Slot,
+    // AMFPropertyStorage
+    pub set_property: Slot,
+    pub get_property: Slot,
+    pub has_property: Slot,
+    pub get_property_count: Slot,
+    pub get_property_at: Slot,
+    pub clear: Slot,
+    pub add_to: Slot,
+    pub copy_to: Slot,
+    pub add_observer: Slot,
+    pub remove_observer: Slot,
+    // AMFContext
+    pub terminate: unsafe extern "system" fn(*mut AmfContext) -> AmfResult,
+    pub init_dx9: Slot,
+    pub get_dx9_device: Slot,
+    pub lock_dx9: Slot,
+    pub unlock_dx9: Slot,
+    pub init_dx11: unsafe extern "system" fn(*mut AmfContext, *mut c_void, i32) -> AmfResult,
+    pub get_dx11_device: Slot,
+    pub lock_dx11: Slot,
+    pub unlock_dx11: Slot,
+    pub init_opencl: Slot,
+    pub get_opencl_context: Slot,
+    pub get_opencl_command_queue: Slot,
+    pub get_opencl_device_id: Slot,
+    pub get_opencl_compute_factory: Slot,
+    pub init_opencl_ex: Slot,
+    pub lock_opencl: Slot,
+    pub unlock_opencl: Slot,
+    pub init_opengl: Slot,
+    pub get_opengl_context: Slot,
+    pub get_opengl_drawable: Slot,
+    pub lock_opengl: Slot,
+    pub unlock_opengl: Slot,
+    pub init_xv: Slot,
+    pub get_xv_device: Slot,
+    pub lock_xv: Slot,
+    pub unlock_xv: Slot,
+    pub init_gralloc: Slot,
+    pub get_gralloc_device: Slot,
+    pub lock_gralloc: Slot,
+    pub unlock_gralloc: Slot,
+    pub alloc_buffer: unsafe extern "system" fn(
+        *mut AmfContext,
+        i32, // AMF_MEMORY_TYPE
+        usize,
+        *mut *mut AmfBuffer,
+    ) -> AmfResult,
+    pub alloc_surface: Slot,
+    pub alloc_audio_buffer: Slot,
+    pub create_buffer_from_host_native: Slot,
+    pub create_surface_from_host_native: Slot,
+    pub create_surface_from_dx9_native: Slot,
+    /// Out-param is `AMFSurface**` in the header; declared as the `AmfData` base here because
+    /// every surface call this module makes (`SetPts`, `SetProperty`, `Release`,
+    /// `SubmitInput`) lives in the `AMFData` vtable prefix, which `AMFSurfaceVtbl` reproduces
+    /// slot-for-slot (single inheritance, same object pointer).
+    pub create_surface_from_dx11_native: unsafe extern "system" fn(
+        *mut AmfContext,
+        *mut c_void,
+        *mut *mut AmfData,
+        *mut c_void,
+    ) -> AmfResult,
+    pub create_surface_from_opengl_native: Slot,
+    pub create_surface_from_gralloc_native: Slot,
+    pub create_surface_from_opencl_native: Slot,
+    pub create_buffer_from_opencl_native: Slot,
+    pub get_compute: Slot,
+}
+
+// -- AMFComponent (components/Component.h) --------------------------------------------------
+#[repr(C)]
+pub struct AmfComponent {
+    pub vtbl: *const AmfComponentVtbl,
+}
+#[repr(C)]
+pub struct AmfComponentVtbl {
+    // AMFInterface
+    pub acquire: Slot,
+    pub release: unsafe extern "system" fn(*mut AmfComponent) -> i32,
+    pub query_interface: Slot,
+    // AMFPropertyStorage
+    pub set_property:
+        unsafe extern "system" fn(*mut AmfComponent, *const u16, AmfVariant) -> AmfResult,
+    pub get_property: Slot,
+    pub has_property: Slot,
+    pub get_property_count: Slot,
+    pub get_property_at: Slot,
+    pub clear: Slot,
+    pub add_to: Slot,
+    pub copy_to: Slot,
+    pub add_observer: Slot,
+    pub remove_observer: Slot,
+    // AMFPropertyStorageEx
+    pub get_properties_info_count: Slot,
+    pub get_property_info_at: Slot,
+    pub get_property_info: Slot,
+    pub validate_property: Slot,
+    // AMFComponent
+    pub init: unsafe extern "system" fn(*mut AmfComponent, i32, i32, i32) -> AmfResult,
+    pub reinit: Slot,
+    pub terminate: unsafe extern "system" fn(*mut AmfComponent) -> AmfResult,
+    pub drain: unsafe extern "system" fn(*mut AmfComponent) -> AmfResult,
+    pub flush: unsafe extern "system" fn(*mut AmfComponent) -> AmfResult,
+    pub submit_input: unsafe extern "system" fn(*mut AmfComponent, *mut AmfData) -> AmfResult,
+    pub query_output: unsafe extern "system" fn(*mut AmfComponent, *mut *mut AmfData) -> AmfResult,
+    pub get_context: Slot,
+    pub set_output_data_allocator_cb: Slot,
+    pub get_caps: Slot,
+    pub optimize: Slot,
+}
+
+// -- AMFData (core/Data.h) — also the usable prefix of AMFSurface --------------------------
+#[repr(C)]
+pub struct AmfData {
+    pub vtbl: *const AmfDataVtbl,
+}
+#[repr(C)]
+pub struct AmfDataVtbl {
+    // AMFInterface
+    pub acquire: Slot,
+    pub release: unsafe extern "system" fn(*mut AmfData) -> i32,
+    pub query_interface:
+        unsafe extern "system" fn(*mut AmfData, *const AmfGuid, *mut *mut c_void) -> AmfResult,
+    // AMFPropertyStorage
+    pub set_property: unsafe extern "system" fn(*mut AmfData, *const u16, AmfVariant) -> AmfResult,
+    pub get_property:
+        unsafe extern "system" fn(*mut AmfData, *const u16, *mut AmfVariant) -> AmfResult,
+    pub has_property: Slot,
+    pub get_property_count: Slot,
+    pub get_property_at: Slot,
+    pub clear: Slot,
+    pub add_to: Slot,
+    pub copy_to: Slot,
+    pub add_observer: Slot,
+    pub remove_observer: Slot,
+    // AMFData
+    pub get_memory_type: Slot,
+    pub duplicate: Slot,
+    pub convert: Slot,
+    pub interop: Slot,
+    pub get_data_type: Slot,
+    pub is_reusable: Slot,
+    pub set_pts: unsafe extern "system" fn(*mut AmfData, i64),
+    pub get_pts: Slot,
+    pub set_duration: Slot,
+    pub get_duration: Slot,
+}
+
+// -- AMFBuffer (core/Buffer.h) — the encoder's output object -------------------------------
+#[repr(C)]
+pub struct AmfBuffer {
+    pub vtbl: *const AmfBufferVtbl,
+}
+#[repr(C)]
+pub struct AmfBufferVtbl {
+    // AMFInterface + AMFPropertyStorage + AMFData prefix (identical order to AmfDataVtbl).
+    pub acquire: Slot,
+    pub release: unsafe extern "system" fn(*mut AmfBuffer) -> i32,
+    pub query_interface: Slot,
+    pub set_property: Slot,
+    pub get_property: Slot,
+    pub has_property: Slot,
+    pub get_property_count: Slot,
+    pub get_property_at: Slot,
+    pub clear: Slot,
+    pub add_to: Slot,
+    pub copy_to: Slot,
+    pub add_observer: Slot,
+    pub remove_observer: Slot,
+    pub get_memory_type: Slot,
+    pub duplicate: Slot,
+    pub convert: Slot,
+    pub interop: Slot,
+    pub get_data_type: Slot,
+    pub is_reusable: Slot,
+    pub set_pts: Slot,
+    pub get_pts: Slot,
+    pub set_duration: Slot,
+    pub get_duration: Slot,
+    // AMFBuffer
+    pub set_size: Slot,
+    pub get_size: unsafe extern "system" fn(*mut AmfBuffer) -> usize,
+    pub get_native: unsafe extern "system" fn(*mut AmfBuffer) -> *mut c_void,
+    pub add_observer_buffer: Slot,
+    pub remove_observer_buffer: Slot,
+}
+
+// -- DLL entry points (core/Factory.h; AMF_CDECL_CALL) --------------------------------------
+pub type AmfQueryVersionFn = unsafe extern "C" fn(*mut u64) -> AmfResult;
+pub type AmfInitFn = unsafe extern "C" fn(u64, *mut *mut AmfFactory) -> AmfResult;
@@ -95,6 +95,13 @@ struct AVD3D11VAFramesContext {
 /// AMD AMF vs Intel QSV — the two libavcodec vendor backends this module covers.
 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
 pub enum WinVendor {
+    /// Benchmark-only, as the module header explains: native AMF replaced the libavcodec AMF path
+    /// in production, and the only remaining CONSTRUCTOR is the `#[cfg(feature = "amf-qsv")]`
+    /// latency A/B in `amf.rs` — the measurement that justifies the native backend existing. That
+    /// is test code, so the *lib* target constructs it nowhere and `dead_code` fires on it (the
+    /// crate root no longer blanket-allows that). Kept deliberately rather than deleted; the arms
+    /// below are what the benchmark drives.
+    #[allow(dead_code)]
    Amf,
    Qsv,
 }
@@ -127,6 +134,44 @@ fn zerocopy_enabled(vendor: WinVendor) -> bool {
        .unwrap_or(matches!(vendor, WinVendor::Amf))
 }

+/// Upper bound on `PUNKTFUNK_FFWIN_POLL_MS`. This knob spins the **encode thread** waiting for an
+/// AU, so a value past one frame period is already self-defeating and a full second is far beyond
+/// anything an operator would set on purpose. The clamp is also what makes the µs conversion
+/// below provably overflow-free.
+///
+/// The reachable hazard is a slipped digit, not the overflow: pre-clamp, `PUNKTFUNK_FFWIN_POLL_MS=
+/// 100000000` was a **27.7-hour** spin with no overflow anywhere near it.
+const MAX_POLL_SPIN_MS: u64 = 1_000;
+
+/// Bounded post-submit spin for [`FfmpegWinEncoder::poll`], in microseconds (0 = off, the default
+/// and the correct choice on every VCN measured so far).
+///
+/// Read from the environment **once per process** (WP6.1): `poll` runs once per encode tick, and
+/// this was an unconditional `env::var` + parse on it.
+///
+/// ⚠ The audit proposed `saturating_mul` for the µs conversion. It is still the wrong fix, but for
+/// a reason worth stating precisely, because the obvious one is false: `Duration::from_micros(
+/// u64::MAX)` is only ~1.8e13 seconds, six orders of magnitude below `Duration`'s `u64::MAX`-second
+/// ceiling, so `Instant::now() + Duration::from_micros(u64::MAX)` does **not** overflow and does
+/// **not** panic (measured, both with and without debug assertions). What it does instead is set a
+/// deadline ~584,000 years out, and the loop below only exits on `Packet`/`Eof` — and this
+/// function's own doc explains that a spin here *provably never* produces the owed AU on the
+/// measured hardware. So `saturating_mul` converts a bad value into a **permanently wedged encode
+/// thread**: a hang, not a panic. Clamping the parsed value first removes the bad value entirely.
+///
+/// (For the record on the pre-clamp behaviour: the workspace sets no `overflow-checks` in
+/// `[profile.release]`, so `ms * 1000` wrapped silently in release and panicked only in debug.)
+fn poll_spin_cap_us() -> u64 {
+    static CAP_US: std::sync::OnceLock<u64> = std::sync::OnceLock::new();
+    *CAP_US.get_or_init(|| {
+        std::env::var("PUNKTFUNK_FFWIN_POLL_MS")
+            .ok()
+            .and_then(|s| s.trim().parse::<u64>().ok())
+            .map(|ms| ms.min(MAX_POLL_SPIN_MS) * 1000)
+            .unwrap_or(0) // default: no spin — the libavcodec AMF buffer can't be spun out
+    })
+}
+
 /// The swscale *source* pixel format for a captured packed-RGB/BGR layout (8-bit BGRA fallback only).
 fn sws_src(format: PixelFormat) -> Result<Pixel> {
    Ok(match format {
@@ -151,8 +196,14 @@ fn sws_src(format: PixelFormat) -> Result<Pixel> {
 }

 /// Does this captured format imply a 10-bit encode (P010 / Rgb10a2)?
-fn is_10bit_format(format: PixelFormat, bit_depth: u8) -> bool {
-    bit_depth >= 10 || matches!(format, PixelFormat::P010 | PixelFormat::Rgb10a2)
+///
+/// Depth follows the PIXELS, not the negotiated `bit_depth` — see
+/// [`crate::ten_bit_input`] for why, and for the failure this shape used to produce here in
+/// particular: a 10-bit-negotiated session over an 8-bit capture built a P010 encoder whose every
+/// `submit_d3d11` then failed the depth check below, forever, with `reset()` unable to help
+/// because the rebuild re-derived the same wrong answer.
+fn is_10bit_format(format: PixelFormat) -> bool {
+    matches!(format, PixelFormat::P010 | PixelFormat::Rgb10a2)
 }

 /// Build the FFmpeg encoder context shared by both inner paths: name, mode, low-latency RC,
@@ -271,6 +322,11 @@ pub fn probe_can_encode_444(_vendor: WinVendor, _codec: Codec) -> bool {
    false
 }

+/// Gated to the builds that can actually reach it: `lib.rs`'s only caller sits under
+/// `cfg(all(not(feature = "qsv"), feature = "amf-qsv"))`, because with the native VPL backend
+/// compiled in it is `qsv::probe_can_encode` that answers. So in the SHIPPED Windows combo
+/// (`nvenc,amf-qsv,qsv`) this function has no caller at all.
+#[cfg(not(feature = "qsv"))]
 pub fn probe_can_encode(vendor: WinVendor, codec: Codec) -> bool {
    // Deliberately NOT pinned to the selected render adapter (unlike `nvenc::probe_can_encode_444`):
    // the system-input probe passes no hwdevice, and the AMF/QSV runtimes only ever bind their own
@@ -350,7 +406,7 @@ impl SystemInner {
        bitrate_bps: u64,
        bit_depth: u8,
    ) -> Result<Self> {
-        let ten_bit = is_10bit_format(format, bit_depth);
+        let ten_bit = crate::ten_bit_input(format, bit_depth);
        let sw_av = if ten_bit {
            ffi::AVPixelFormat::AV_PIX_FMT_P010LE
        } else {
@@ -472,7 +528,10 @@ impl SystemInner {
        pts: i64,
        idr: bool,
    ) -> Result<()> {
-        let fmt_10 = matches!(format, PixelFormat::P010 | PixelFormat::Rgb10a2);
+        // Same predicate the encoder was built from (`ten_bit_input`), so this can only fire on a
+        // genuine MID-STREAM depth change — never, as it used to, on every frame of a session that
+        // merely negotiated 10-bit over an 8-bit capture.
+        let fmt_10 = is_10bit_format(format);
        anyhow::ensure!(
            fmt_10 == self.ten_bit,
            "captured format {format:?} bit-depth changed under the encoder (built {}-bit)",
@@ -851,7 +910,7 @@ impl ZeroCopyInner {
        bit_depth: u8,
        device: &ID3D11Device,
    ) -> Result<Self> {
-        let ten_bit = is_10bit_format(format, bit_depth);
+        let ten_bit = crate::ten_bit_input(format, bit_depth);
        let sw_av = if ten_bit {
            ffi::AVPixelFormat::AV_PIX_FMT_P010LE
        } else {
@@ -1311,11 +1370,7 @@ impl Encoder for FfmpegWinEncoder {
            Some(Inner::ZeroCopy(z)) => &mut z.enc,
            None => return Ok(None),
        };
-        let cap_us = std::env::var("PUNKTFUNK_FFWIN_POLL_MS")
-            .ok()
-            .and_then(|s| s.parse::<u64>().ok())
-            .map(|ms| ms * 1000)
-            .unwrap_or(0); // default: no spin — the libavcodec AMF buffer can't be spun out
+        let cap_us = poll_spin_cap_us();
        let deadline = (cap_us > 0 && self.in_flight > 0)
            .then(|| std::time::Instant::now() + std::time::Duration::from_micros(cap_us));
        loop {
@@ -37,8 +37,9 @@
 #![deny(clippy::undocumented_unsafe_blocks)]

 use super::nvenc_core::{
-    apply_low_latency_config, build_init_params, codec_guid, resolve_slices, resolve_subframe,
-    LowLatencyConfig, NvStatusExt, RFI_DPB,
+    apply_low_latency_config, build_init_params, cached_ceiling, codec_guid, plan_range_recovery,
+    resolve_slices, resolve_split_mode, resolve_subframe, store_ceiling, CeilingKey,
+    LowLatencyConfig, NvStatusExt, RangePlan,
 };
 use super::nvenc_status;
 use super::{ChromaFormat, Codec, EncodedFrame, Encoder, EncoderCaps};
@@ -263,6 +264,112 @@ fn split_mode_units(split_mode: u32) -> u32 {
    }
 }

+/// Serializes every `nvEncOpenEncodeSessionEx` in this process against [`reap_parked_sessions`].
+/// The reap retry-destroys handles whose driver-side state is unknown; if it overlapped an open,
+/// the driver could hand the NEW session the recycled address of the zombie being destroyed and
+/// the reap would destroy a live session. Held across `init_session`'s whole open sequence
+/// (including its caps probe and bitrate-clamp search) and around the standalone caps probes.
+/// Admission (`can_open_another_session`) never takes it — that stays a lock-free atomic load.
+static DRIVER_SESSION_GATE: std::sync::Mutex<()> = std::sync::Mutex::new(());
+
+/// A session whose `destroy_encoder` failed with an AMBIGUOUS status (see
+/// [`nvenc_status::destroy_proves_no_session`]): the driver may still hold its concurrent-session
+/// slot, so its units stay charged in [`LIVE_SESSION_UNITS`] (fail-closed for admission) and the
+/// handle is parked here for [`reap_parked_sessions`] to retry. The retry is what keeps a
+/// TRANSIENT failure (wedge episode a TDR later cleared) from poisoning the budget until a host
+/// restart, while a genuine driver leak keeps failing the retry and stays correctly charged.
+struct ParkedSession {
+    enc: usize,
+    units: u32,
+    /// Pins the D3D11 device the session was opened against. Teardown drops the texture
+    /// registrations (the only other refs), and the reinit-on-device-change path parks exactly
+    /// when the old device is on its way out — without this, the reap's late `destroy_encoder`
+    /// could touch a freed device.
+    _device: Option<ID3D11Device>,
+}
+
+// SAFETY: the COM ref is the only non-Send field. The D3D11 device is free-threaded (capture
+// creates it without `D3D11_CREATE_DEVICE_SINGLETHREADED` — see pf-frame/src/dxgi.rs), we never
+// call methods on it from here, and dropping (Release) a free-threaded COM object from another
+// thread is sound. The raw `enc` handle is only ever passed back to `destroy_encoder` under
+// [`PARKED`]'s lock with [`DRIVER_SESSION_GATE`] held — single-threaded access in practice.
+unsafe impl Send for ParkedSession {}
+
+static PARKED: std::sync::Mutex<Vec<ParkedSession>> = std::sync::Mutex::new(Vec::new());
+
+/// Park a session whose destroy failed ambiguously. Units are NOT refunded — they keep counting
+/// against admission until a reap proves the slot free. Bounded: once parked units would exceed
+/// the session cap the oldest entry ages out WITH a refund (a repeated wedge-rebuild loop would
+/// otherwise grow this without limit, and by then the driver has almost certainly been through
+/// the reset that reclaims the early handles anyway).
+fn park_session(enc: usize, units: u32, device: Option<ID3D11Device>) {
+    let mut parked = PARKED.lock().unwrap_or_else(|p| p.into_inner());
+    while !parked.is_empty() && parked.iter().map(|z| z.units).sum::<u32>() + units > session_cap()
+    {
+        let old = parked.remove(0);
+        LIVE_SESSION_UNITS.fetch_sub(old.units, std::sync::atomic::Ordering::Relaxed);
+        tracing::warn!(
+            enc = old.enc,
+            units = old.units,
+            "NVENC parked-session graveyard full — aging out the oldest entry (units refunded)"
+        );
+    }
+    parked.push(ParkedSession {
+        enc,
+        units,
+        _device: device,
+    });
+}
+
+/// Retry `destroy_encoder` on every parked session and refund the units of each one that now
+/// succeeds (or fails with a session-gone status — same proof). Runs from `init_session` on the
+/// encode thread, NEVER from admission (a wedged driver would block the admission registry lock).
+///
+/// # Safety
+/// Caller must hold [`DRIVER_SESSION_GATE`] and there must be NO live NVENC session in the
+/// process (checked here: live units == parked units). Under those two conditions no live session
+/// can alias a recycled handle address, and no concurrent open can be handed one mid-reap. The
+/// residual assumption — documented, unprovable from the SDK — is that a FAILED destroy leaves
+/// the handle itself intact for a later retry; NVENC documents no retry semantics either way.
+unsafe fn reap_parked_sessions() {
+    let mut parked = PARKED.lock().unwrap_or_else(|p| p.into_inner());
+    if parked.is_empty() {
+        return;
+    }
+    let parked_units: u32 = parked.iter().map(|z| z.units).sum();
+    if LIVE_SESSION_UNITS.load(std::sync::atomic::Ordering::Relaxed) != parked_units {
+        return; // a live session exists somewhere — its address space is off limits
+    }
+    parked.retain(
+        |z| match (api().destroy_encoder)(z.enc as *mut c_void).nv_ok() {
+            Ok(()) => {
+                tracing::info!(
+                    enc = z.enc,
+                    units = z.units,
+                    "NVENC parked session reclaimed — retry-destroy succeeded, budget refunded"
+                );
+                LIVE_SESSION_UNITS.fetch_sub(z.units, std::sync::atomic::Ordering::Relaxed);
+                false
+            }
+            Err(e) if nvenc_status::destroy_proves_no_session(e) => {
+                tracing::info!(
+                    enc = z.enc,
+                    units = z.units,
+                    status = ?e,
+                    "NVENC parked session gone on the driver side — budget refunded"
+                );
+                LIVE_SESSION_UNITS.fetch_sub(z.units, std::sync::atomic::Ordering::Relaxed);
+                false
+            }
+            Err(e) => {
+                tracing::debug!(enc = z.enc, status = ?e,
+                    "NVENC parked session still refuses destroy — units stay charged");
+                true
+            }
+        },
+    );
+}
+
 /// Whether the operator asked for the two-thread async retrieve (`PUNKTFUNK_NVENC_ASYNC` truthy).
 /// Combined with the GPU's `NV_ENC_CAPS_ASYNC_ENCODE_SUPPORT` in `init_session`. Opt-in until
 /// on-glass validated; note an async-rejecting config surfaces as a failed session open — unset
@@ -279,12 +386,24 @@ fn async_retrieve_requested() -> bool {
 /// the ring textures in place, so in-flight depth beyond the ring lets the capturer overwrite a
 /// frame mid-encode: visual corruption, not UB). IDD-push rings are sized around
 /// `PUNKTFUNK_IDD_DEPTH`; raise both together if deeper pipelining is needed.
+///
+/// Read from the environment **once per process** (WP6.1): `submit` consults this on EVERY frame —
+/// both arms of the `cap` match, in sync mode too, where the result is then discarded because the
+/// backpressure loop short-circuits on `async_rt`. Memoizing rather than latching into a session
+/// field is deliberate: the composed `cap` also folds in `input_ring_depth`, which
+/// `set_input_ring_depth` may change after open, and freezing that half would reintroduce the
+/// in-place-overwrite bug the ring term exists to prevent. Nothing in the workspace mutates this
+/// variable at runtime, so memoizing a process-constant read is behaviour-preserving by
+/// construction.
 fn async_inflight_cap() -> usize {
-    std::env::var("PUNKTFUNK_NVENC_ASYNC_DEPTH")
-        .ok()
-        .and_then(|s| s.parse::<usize>().ok())
-        .unwrap_or(4)
-        .clamp(2, POOL - 1)
+    static CAP: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
+    *CAP.get_or_init(|| {
+        std::env::var("PUNKTFUNK_NVENC_ASYNC_DEPTH")
+            .ok()
+            .and_then(|s| s.trim().parse::<usize>().ok())
+            .unwrap_or(4)
+            .clamp(2, POOL - 1)
+    })
 }

 /// One in-flight encode handed to the retrieve thread: the output bitstream to lock once its
@@ -323,7 +442,23 @@ fn retrieve_loop(
    done_tx: mpsc::Sender<RetrieveDone>,
 ) {
    pf_frame::thread_qos::boost_thread_priority(false);
+    // Wedged-vs-routine drain (WP5.2b). Teardown drops the job sender and joins this thread, so
+    // every queued job is WAITED for (never abandoned — abandoning converts the routine drain
+    // into destroy/unmap/close-while-encoding). But on a wedged session that used to cost the
+    // full 5 s PER queued job: up to `cap x 5 s` on the encode thread, and the stall-recovery
+    // ladder rebuilds the session several times per episode, paying it each rebuild. The wedge
+    // evidence lives right here — a completion wait that burned its full budget — so after one
+    // full-budget timeout later jobs get a short slice instead. A first timeout is already
+    // encoder-fatal (its Err reaches `absorb_done` before any later completion is handed out,
+    // and the host tears the session down), so short-slicing behind an established wedge changes
+    // only how long teardown blocks, never an outcome. A successful wait resets the latch, which
+    // keeps a transient GPU stall from cascading short slices into false timeouts. (NVENC does
+    // not document completion ordering across in-flight frames; nothing here relies on it —
+    // FIFO only sharpens the latency bound.)
+    const WEDGED_DRAIN_WAIT_MS: u32 = 250;
+    let mut wedged = false;
    while let Ok(job) = work_rx.recv() {
+        let wait_ms = if wedged { WEDGED_DRAIN_WAIT_MS } else { 5000 };
        // SAFETY: `job.event` is one of the auto-reset events `init_session` created and
        // registered for exactly this session, and `job.bs` one of its pool bitstreams; both stay
        // valid until `teardown`, which joins this thread first. `WaitForSingleObject` takes the
@@ -334,9 +469,13 @@ fn retrieve_loop(
        // Lock/unlock from a secondary thread while the encode thread submits is the NVENC
        // guide's documented threading model.
        let result = unsafe {
-            if WaitForSingleObject(HANDLE(job.event as *mut c_void), 5000) != WAIT_OBJECT_0 {
-                Err("NVENC completion event timeout (5s) — encoder wedged?".to_string())
+            if WaitForSingleObject(HANDLE(job.event as *mut c_void), wait_ms) != WAIT_OBJECT_0 {
+                wedged = true;
+                Err(format!(
+                    "NVENC completion event timeout ({wait_ms} ms) — encoder wedged?"
+                ))
            } else {
+                wedged = false;
                let mut lock = nv::NV_ENC_LOCK_BITSTREAM {
                    version: nv::NV_ENC_LOCK_BITSTREAM_VER,
                    outputBitstream: job.bs as *mut c_void,
@@ -387,6 +526,12 @@ pub struct NvencD3d11Encoder {
    /// `NV_ENC_CAPS_SUPPORT_YUV444_ENCODE` (cleared in `query_caps` on a card that lacks it) and on an
    /// RGB input format (NV12/P010 capture can't reconstruct 4:4:4). HEVC-only.
    chroma_444: bool,
+    /// What the SESSION NEGOTIATED, before any per-init downgrade. `chroma_444` above is the
+    /// EFFECTIVE value for the session currently open and is cleared whenever the capturer hands us
+    /// subsampled YUV — which used to overwrite the negotiation itself, so a capturer that went
+    /// back to RGB (HDR toggle, capture-path switch) could never recover 4:4:4 for the rest of the
+    /// stream. Keeping the request separate lets every re-init recompute the effective value.
+    chroma_444_requested: bool,
    /// `NV_ENC_CAPS_SUPPORT_YUV444_ENCODE` from the caps probe — whether this GPU can 4:4:4 encode at
    /// all. `chroma_444` is forced off when this is false (graceful downgrade to 4:2:0).
    yuv444_supported: bool,
@@ -394,6 +539,17 @@ pub struct NvencD3d11Encoder {
    /// `ABGR10` input format + the BT.2020/PQ colour VUI. Derived per-frame from the capture format
    /// (HDR can toggle mid-session); a change re-inits the session.
    hdr: bool,
+    /// The HDR state the CAPTURE FORMAT asks for, independent of whether this GPU can serve it.
+    ///
+    /// `hdr` above is the effective value and `query_caps` clears it on a card without 10-bit
+    /// encode. `submit` re-derives the requested value from every frame's pixel format, so
+    /// comparing that against the post-downgrade `hdr` made them disagree permanently: a P010
+    /// capturer on such a GPU reported "HDR changed" on EVERY frame and tore down and rebuilt the
+    /// whole session each time. The re-init trigger compares against THIS instead.
+    hdr_requested: bool,
+    /// Latched when `query_caps` finds no 10-bit encode support, so the downgrade is remembered
+    /// across re-inits instead of being rediscovered (and re-warned) every session.
+    hdr_unsupported: bool,
    /// The source's static HDR mastering metadata (from the capturer's `GetDesc1`), emitted as
    /// in-band SEI (`mastering_display_colour_volume` + `content_light_level_info`) on each keyframe
    /// when `hdr`. `None` = unknown → no SEI (the VUI still signals BT.2020 PQ). Set per-frame via
@@ -455,6 +611,11 @@ pub struct NvencD3d11Encoder {
    /// device on a desktop switch (normal ↔ Winlogon secure); when a frame carries a new device we
    /// tear down and re-init NVENC against it.
    init_device: *mut c_void,
+    /// COM ref pinning that device while the session lives. `init_device` alone pins nothing, and
+    /// `teardown` releases the texture registrations (the only other refs) BEFORE destroying the
+    /// session — worse, a failed destroy parks the session handle for a later retry, which must
+    /// never outlive the device it references. Moved into the [`ParkedSession`] on that path.
+    init_device_com: Option<ID3D11Device>,
    /// The hardware-session units THIS encoder holds against [`LIVE_SESSION_UNITS`] (1 plain, 2–3
    /// under forced split-encode — a split session occupies one session per engine). `0` while no
    /// session is open; set by `init_session`, returned by `teardown`.
@@ -504,6 +665,9 @@ impl NvencD3d11Encoder {
            bit_depth,
            // 4:4:4 is HEVC-only; the GPU-support gate is applied in `query_caps`.
            chroma_444: chroma.is_444() && codec == Codec::H265,
+            chroma_444_requested: chroma.is_444() && codec == Codec::H265,
+            hdr_requested: false,
+            hdr_unsupported: false,
            yuv444_supported: false,
            hdr: false,
            hdr_meta: None,
@@ -525,6 +689,7 @@ impl NvencD3d11Encoder {
            session_async: false,
            last_rfi_range: None,
            init_device: ptr::null_mut(),
+            init_device_com: None,
            session_units: 0,
        })
    }
@@ -571,18 +736,41 @@ impl NvencD3d11Encoder {
        for &bs in &self.bitstreams {
            let _ = (api().destroy_bitstream_buffer)(self.encoder, bs);
        }
-        // A destroy failure means the driver may still hold this session's slot (the concurrent-
-        // session cap is per process and only a restart clears a leak) — make it visible instead
-        // of silently discarding the status.
-        if let Err(e) = (api().destroy_encoder)(self.encoder).nv_ok() {
-            tracing::warn!(
-                status = ?e,
-                "NVENC destroy_encoder failed at teardown — the driver may have leaked this \
-                 session's slot toward the concurrent-session cap"
-            );
+        // Session-budget truth (see LIVE_SESSION_UNITS): refund only on PROOF the driver released
+        // the slot — a successful destroy, or a failure whose status says the session/device no
+        // longer exists on the driver side (a TDR reclaims every session with the context). The
+        // old unconditional refund made the counter drift low on real leaks (over-admitting
+        // parallel displays); the opposite extreme — a permanent charge on ANY failure — let one
+        // transient wedge episode poison admission until a host restart. Ambiguous failures
+        // instead PARK the handle with its units still charged (fail-closed), and
+        // `reap_parked_sessions` retries the destroy once no session is live, so the charge lasts
+        // exactly as long as the driver keeps refusing — which is the definition of still-leaked.
+        let dev_pin = self.init_device_com.take();
+        match (api().destroy_encoder)(self.encoder).nv_ok() {
+            Ok(()) => {
+                LIVE_SESSION_UNITS
+                    .fetch_sub(self.session_units, std::sync::atomic::Ordering::Relaxed);
+            }
+            Err(e) if nvenc_status::destroy_proves_no_session(e) => {
+                tracing::warn!(
+                    status = ?e,
+                    "NVENC destroy_encoder failed, but the status proves the driver holds no \
+                     session (device gone/reset) — budget refunded"
+                );
+                LIVE_SESSION_UNITS
+                    .fetch_sub(self.session_units, std::sync::atomic::Ordering::Relaxed);
+            }
+            Err(e) => {
+                tracing::warn!(
+                    status = ?e,
+                    units = self.session_units,
+                    "NVENC destroy_encoder failed ambiguously — the driver may still hold this \
+                     session's slot; parking the handle (units stay charged until a reap-destroy \
+                     proves the slot free)"
+                );
+                park_session(self.encoder as usize, self.session_units, dev_pin);
+            }
        }
-        // Return this session's units to the budget (see LIVE_SESSION_UNITS).
-        LIVE_SESSION_UNITS.fetch_sub(self.session_units, std::sync::atomic::Ordering::Relaxed);
        self.session_units = 0;
        self.regs.clear(); // drops the texture clones, releasing our refs
        self.bitstreams.clear();
@@ -641,6 +829,9 @@ impl NvencD3d11Encoder {
                e,
            ));
        }
+        // The handshake with the kernel-mode driver just succeeded — from here on, an
+        // `NV_ENC_ERR_INVALID_VERSION` in this process cannot be a driver version skew.
+        nvenc_status::note_session_opened();
        let wmax = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_WIDTH_MAX);
        let hmax = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_HEIGHT_MAX);
        let ten_bit = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_10BIT_ENCODE);
@@ -668,7 +859,12 @@ impl NvencD3d11Encoder {
        }
        // Degrade gracefully rather than fail: no 10-bit encode on this card → 8-bit SDR.
        if self.bit_depth >= 10 && ten_bit == 0 {
-            tracing::warn!("NVENC: this GPU can't 10-bit encode — falling back to 8-bit SDR");
+            if !self.hdr_unsupported {
+                tracing::warn!("NVENC: this GPU can't 10-bit encode — falling back to 8-bit SDR");
+            }
+            // Latched: `submit` compares the frame's REQUESTED hdr against `hdr_requested`, not
+            // against this cleared value, so a 10-bit capturer no longer re-inits every frame.
+            self.hdr_unsupported = true;
            self.bit_depth = 8;
            self.hdr = false;
        }
@@ -754,6 +950,27 @@ impl NvencD3d11Encoder {
        Ok(cfg)
    }

+    /// This session config's identity in the process-lifetime bitrate-ceiling cache
+    /// (`nvenc_core::{cached_ceiling, store_ceiling}`). GPU identity is the selected render
+    /// adapter's LUID — the adapter the capturer's device (and so this session) lives on; `0`
+    /// when unresolved. Best effort by design: the cache is advisory, a colliding identity costs
+    /// one failed open + re-search, never a wrong session.
+    fn ceiling_key(&self, split_mode: u32) -> CeilingKey {
+        let gpu = pf_gpu::resolve_render_adapter_luid()
+            .map(|l| ((l.HighPart as u32 as u64) << 32) | l.LowPart as u64)
+            .unwrap_or(0);
+        CeilingKey {
+            gpu,
+            codec: self.codec,
+            width: self.width,
+            height: self.height,
+            fps: self.fps,
+            bit_depth: self.bit_depth,
+            chroma_444: self.chroma_444,
+            split_mode,
+        }
+    }
+
    /// Open + configure + initialize ONE NVENC session at `bitrate` (bps) and `split_mode`. Returns
    /// the session handle, or destroys it and returns the error. NVENC has no re-init after a failed
    /// `initialize_encoder`, so the bitrate-clamp search in `init_session` calls this once per probe.
@@ -780,6 +997,7 @@ impl NvencD3d11Encoder {
            }
            return Err(nvenc_status::call_err("open_encode_session_ex", e));
        }
+        nvenc_status::note_session_opened();

        let mut cfg = match self.build_config(enc, bitrate) {
            Ok(cfg) => cfg,
@@ -812,6 +1030,15 @@ impl NvencD3d11Encoder {

    /// Lazily create the session on the first frame's D3D11 device (so capture + encode share it).
    fn init_session(&mut self, device: &ID3D11Device) -> Result<()> {
+        // Serialize this whole open sequence (caps probe + bitrate-clamp search + charge) against
+        // every other open and against the zombie reap — see DRIVER_SESSION_GATE. Cold path:
+        // sessions open on a stream start or a device-change rebuild, never per frame.
+        let _gate = DRIVER_SESSION_GATE
+            .lock()
+            .unwrap_or_else(|p| p.into_inner());
+        // SAFETY: gate held (no open can be handed a recycled zombie address mid-reap) and the
+        // reap itself re-checks that no live session exists before touching any parked handle.
+        unsafe { reap_parked_sessions() };
        // SAFETY: every call below goes through a function pointer resolved once from the
        // runtime-loaded [`EncodeApi`] table (`api()`, gated in `open`), or through this type's own
        // `unsafe fn`s whose contract is met here. `query_caps`/`try_open_session` receive `device`,
@@ -834,43 +1061,14 @@ impl NvencD3d11Encoder {
            // gets the highest the GPU can actually do, not a coarse fraction of it.
            const FLOOR_BPS: u64 = 10_000_000;
            let requested_bps = self.bitrate_bps;
-            // 2-way NVENC split-frame encoding (Ada dual-NVENC) — the high-pixel-rate throughput lever
-            // the Linux host enables via libavcodec `split_encode_mode`. A single Ada NVENC session tops
-            // out ~0.8 Gpix/s, so at high motion a 5K@240 (1.77 Gpix/s) frame takes ~8 ms to encode and
-            // the rate caps ~125 fps; splitting across both engines roughly halves that. Force 2-way
-            // above ~1 Gpix/s (matching encode/linux.rs), AUTO below (the ~2% BD-rate cost isn't worth
-            // it at low pixel rates). Env override PUNKTFUNK_SPLIT_ENCODE = 0/disable | 1/auto | 2 | 3.
-            // HEVC/AV1 only; the init-failure fallback below disables it if a codec/config rejects it.
+            // 2-way NVENC split-frame encoding (Ada dual-NVENC) — the high-pixel-rate throughput lever.
+            // A single Ada NVENC session tops out ~0.8-1 Gpix/s, so at high motion a 5K@240
+            // (1.77 Gpix/s) frame takes ~8 ms to encode and the rate caps ~125 fps; splitting across
+            // both engines roughly halves that. Shared selector — see [`resolve_split_mode`] for the
+            // precedence (env override / the measured Main10 don't-split rule / pixel rate).
+            // The init-failure fallback below disables it if a codec/config rejects it.
            let pixel_rate = self.width as u64 * self.height as u64 * self.fps.max(1) as u64;
-            let mut split_mode: u32 = match std::env::var("PUNKTFUNK_SPLIT_ENCODE").ok().as_deref()
-            {
-                Some("0") | Some("disable") => {
-                    nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32
-                }
-                Some("1") | Some("auto") => {
-                    nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_AUTO_FORCED_MODE as u32
-                }
-                Some("3") => nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_THREE_FORCED_MODE as u32,
-                Some("2") => nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_TWO_FORCED_MODE as u32,
-                // Main10 (10-bit / HDR): 2-way split is measurably SLOWER on Ada — at 5120x1440@240
-                // Main10, forced-2 took 7.6 ms/frame (~131 fps) vs 2.8 ms (~357 fps) single-engine
-                // (the split/merge overhead dominates for 10-bit). A single Ada NVENC engine already
-                // handles 5K@240 Main10 well under the 4.17 ms budget, so DON'T split — splitting was
-                // the "broken animations in HDR" (the stream capped at ~131 fps). Env still overrides.
-                _ if self.bit_depth >= 10 => {
-                    nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32
-                }
-                _ if pixel_rate > 1_000_000_000 => {
-                    nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_TWO_FORCED_MODE as u32
-                }
-                _ => nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_AUTO_MODE as u32,
-            };
-            tracing::debug!(
-                split_mode,
-                bit_depth = self.bit_depth,
-                pixel_rate,
-                "NVENC split-encode mode selected"
-            );
+            let split_mode: u32 = resolve_split_mode(self.bit_depth, pixel_rate);
            // Find the highest bitrate the GPU's codec LEVEL accepts and CLAMP to it. NVENC rejects
            // `initialize_encoder` (InvalidParam) when the bitrate exceeds the level ceiling (e.g. a
            // 1 Gbps request on HEVC). Strategy: try the requested rate; if the only problem is a forced
@@ -884,39 +1082,69 @@ impl NvencD3d11Encoder {
            // built in the right mode from the start.
            let use_async = self.async_supported && async_retrieve_requested();

-            let mut probe = self.try_open_session(device, requested_bps, split_mode, use_async);
+            // Ceiling cache (process lifetime, `nvenc_core`): a prior clamp search already found
+            // this config's max accepted rate — open straight AT the ceiling instead of paying
+            // the ~6-open binary search (and its session churn) on every ABR overshoot.
+            let mut target_bps = requested_bps;
+            if let Some(ceiling) = cached_ceiling(&self.ceiling_key(split_mode)) {
+                if requested_bps > ceiling {
+                    tracing::info!(
+                        requested_mbps = requested_bps / 1_000_000,
+                        ceiling_mbps = ceiling / 1_000_000,
+                        "NVENC: requested bitrate above the cached codec-level ceiling — opening \
+                         at the ceiling"
+                    );
+                    target_bps = ceiling;
+                }
+            }
+
+            let mut probe = self.try_open_session(device, target_bps, split_mode, use_async);
+            // The cache is advisory: a stale entry (driver change, identity collision) must not
+            // wedge the open — retry the requested rate and let the search below rediscover.
+            if probe.is_err() && target_bps < requested_bps {
+                target_bps = requested_bps;
+                probe = self.try_open_session(device, requested_bps, split_mode, use_async);
+            }
            // Disambiguate a forced-split rejection from a bitrate-cap rejection: retry once at the
            // requested rate with split disabled — if THAT succeeds, split was the problem, not bitrate.
            // ANY non-disabled mode can be the rejection — AUTO included: AV1 rejects the whole
            // init with INVALID_PARAM on drivers/configs where auto split isn't valid for it,
            // which then masqueraded as a bitrate cap and failed "even at the floor".
+            // `used_split` tracks the mode sessions ACTUALLY open with from here on — it feeds
+            // `self.split_mode` (a reconfigure must re-present it) and the ceiling-cache key.
+            let mut used_split = split_mode;
            let split_on =
                split_mode != nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32;
            if probe.is_err() && split_on {
                let no_split = nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32;
-                if let Ok(e) = self.try_open_session(device, requested_bps, no_split, use_async) {
+                if let Ok(e) = self.try_open_session(device, target_bps, no_split, use_async) {
                    tracing::warn!("NVENC: split-encode rejected by codec/config — disabled");
-                    split_mode = no_split;
+                    used_split = no_split;
                    probe = Ok(e);
                }
            }

            let enc = match probe {
                Ok(enc) => {
-                    self.bitrate_bps = requested_bps;
+                    self.bitrate_bps = target_bps;
                    enc
                }
+                // Only a parameter/caps rejection means "the bitrate is above the codec-level
+                // ceiling". A transient failure (busy engine, session limit, OOM, device loss,
+                // version skew) must propagate — a search steered by it would discover, and
+                // cache, a bogus ceiling.
+                Err(e) if !nvenc_status::is_param_rejection(&e) => return Err(e),
                Err(_) => {
                    // Requested bitrate exceeds the codec-level ceiling — binary-search the max accepted.
                    // `lo` is the highest known-good rate (FLOOR is assumed to fit), `hi` the lowest
                    // rejected; `best` holds the live session at `lo` so we end up with the clamped one.
                    let mut lo = FLOOR_BPS;
-                    let mut hi = requested_bps;
+                    let mut hi = target_bps;
                    let mut best: *mut c_void = ptr::null_mut();
                    let mut best_bps = 0u64;
                    while hi > lo + CLAMP_TOL_BPS {
                        let mid = lo + (hi - lo) / 2;
-                        match self.try_open_session(device, mid, split_mode, use_async) {
+                        match self.try_open_session(device, mid, used_split, use_async) {
                            Ok(e) => {
                                if !best.is_null() {
                                    let _ = (api().destroy_encoder)(best);
@@ -925,7 +1153,15 @@ impl NvencD3d11Encoder {
                                best_bps = mid;
                                lo = mid;
                            }
-                            Err(_) => hi = mid,
+                            Err(e) if nvenc_status::is_param_rejection(&e) => hi = mid,
+                            Err(e) => {
+                                // Environmental mid-search failure: don't let it shrink the
+                                // search — release the partial result and propagate.
+                                if !best.is_null() {
+                                    let _ = (api().destroy_encoder)(best);
+                                }
+                                return Err(e);
+                            }
                        }
                    }
                    if best.is_null() {
@@ -933,14 +1169,19 @@ impl NvencD3d11Encoder {
                        // trying split-disabled in case a forced split (not the bitrate) is the blocker.
                        let no_split =
                            nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32;
-                        best = self
-                            .try_open_session(device, FLOOR_BPS, split_mode, use_async)
-                            .or_else(|_| {
-                                self.try_open_session(device, FLOOR_BPS, no_split, use_async)
-                            })
-                            .context(
-                                "NVENC initialize_encoder rejected even at the floor bitrate",
-                            )?;
+                        best = match self.try_open_session(device, FLOOR_BPS, used_split, use_async)
+                        {
+                            Ok(e) => e,
+                            Err(_) => {
+                                let e = self
+                                    .try_open_session(device, FLOOR_BPS, no_split, use_async)
+                                    .context(
+                                    "NVENC initialize_encoder rejected even at the floor bitrate",
+                                )?;
+                                used_split = no_split;
+                                e
+                            }
+                        };
                        best_bps = FLOOR_BPS;
                    }
                    tracing::warn!(
@@ -948,21 +1189,22 @@ impl NvencD3d11Encoder {
                        clamped_mbps = best_bps / 1_000_000,
                        "NVENC: requested bitrate above the GPU codec-level ceiling — clamped to the max accepted"
                    );
+                    store_ceiling(self.ceiling_key(used_split), best_bps);
                    self.bitrate_bps = best_bps;
                    best
                }
            };
            self.encoder = enc;
-            // Session init params a later `reconfigure_bitrate` must re-present verbatim. (Best
-            // effort: the floor fallback above may have succeeded split-disabled without updating
-            // `split_mode` — a reconfigure then presents the forced mode, NVENC rejects it, and
-            // the caller's rebuild fallback covers the mismatch.)
-            self.split_mode = split_mode;
+            // Pin the device for the session's lifetime — and for a parked afterlife if the
+            // eventual destroy fails (see `ParkedSession::_device`).
+            self.init_device_com = Some(device.clone());
+            // Session init params a later `reconfigure_bitrate` must re-present verbatim.
+            self.split_mode = used_split;
            self.session_async = use_async;
            // Session-budget accounting (Stage W3): record what this open holds so admission can
            // decline a parallel display the hardware can't afford. Weighted by the FINAL split
            // mode (a split session occupies one hardware session per engine).
-            self.session_units = split_mode_units(split_mode);
+            self.session_units = split_mode_units(used_split);
            LIVE_SESSION_UNITS.fetch_add(self.session_units, std::sync::atomic::Ordering::Relaxed);
            // (The clamp path above already logs the requested→clamped bitrate at warn; no second
            // info line for the same event here.)
@@ -986,6 +1228,12 @@ impl NvencD3d11Encoder {
                for _ in 0..POOL {
                    let ev = CreateEventW(None, false, false, PCWSTR::null())
                        .context("CreateEvent (NVENC completion)")?;
+                    // Push BEFORE registering: a failed registration propagates out of
+                    // `init_session` into submit's teardown call, and only handles that made it
+                    // into `self.events` get closed there — pushing after the fallible call
+                    // leaked the Win32 event on every registration failure. Teardown's
+                    // unregister of the one never-registered event is a harmless error return.
+                    self.events.push(ev.0 as usize);
                    let mut ep = nv::NV_ENC_EVENT_PARAMS {
                        version: nv::NV_ENC_EVENT_PARAMS_VER,
                        completionEvent: ev.0,
@@ -994,7 +1242,6 @@ impl NvencD3d11Encoder {
                    (api().register_async_event)(enc, &mut ep)
                        .nv_ok()
                        .map_err(|e| nvenc_status::call_err("register_async_event", e))?;
-                    self.events.push(ev.0 as usize);
                }
                let (work_tx, work_rx) = mpsc::sync_channel::<RetrieveJob>(POOL);
                let (done_tx, done_rx) = mpsc::channel::<RetrieveDone>();
@@ -1086,7 +1333,11 @@ impl Encoder for NvencD3d11Encoder {
        let dev_raw = frame.device.as_raw();
        let size_changed =
            self.inited && (self.width != captured.width || self.height != captured.height);
-        let hdr_changed = self.inited && self.hdr != hdr;
+        // Compare against what was REQUESTED last init, not the effective `self.hdr`: on a GPU
+        // without 10-bit encode `query_caps` clears `self.hdr`, so comparing it against a P010
+        // capturer's perpetual `hdr = true` reported a change on every single frame and tore the
+        // session down and rebuilt it each time.
+        let hdr_changed = self.inited && self.hdr_requested != hdr;
        if self.inited && (self.init_device != dev_raw || size_changed || hdr_changed) {
            tracing::info!(
                device_changed = self.init_device != dev_raw,
@@ -1107,7 +1358,13 @@ impl Encoder for NvencD3d11Encoder {
            // Adopt the current frame size + colour so the encoder always matches the capturer output.
            self.width = captured.width;
            self.height = captured.height;
+            self.hdr_requested = hdr;
+            // Effective until `query_caps` says otherwise (it clears this on a card without 10-bit).
            self.hdr = hdr;
+            // Recompute the EFFECTIVE 4:4:4 from the negotiation on every init. This used to read
+            // (and overwrite) `chroma_444` itself, so the first subsampled-YUV frame permanently
+            // demoted the session — 4:4:4 never returned even when the capturer went back to RGB.
+            self.chroma_444 = self.chroma_444_requested;
            // Pick the NVENC input format from the captured pixel format. YUV (NV12/P010) is the
            // video-processor path — NVENC encodes it natively (no internal RGB→YUV, which is a hidden
            // 3D/compute step that would fight a GPU-saturating game). RGB (ARGB/ABGR10) is the legacy
@@ -1133,9 +1390,11 @@ impl Encoder for NvencD3d11Encoder {
            };
            // 4:4:4 honesty: the FREXT/chromaFormatIDC=3 config engages only on an RGB input (a
            // subsampled NV12/P010 source can't reconstruct full chroma). If the capturer handed
-            // native YUV despite a 4:4:4 negotiation, this session encodes 4:2:0 — clear the flag
-            // NOW so `caps().chroma_444` (and native's post-open cross-check) reports what
-            // the stream really carries instead of silently claiming full chroma.
+            // native YUV despite a 4:4:4 negotiation, THIS session encodes 4:2:0 — clear the
+            // effective flag now so `caps().chroma_444` (and native's post-open cross-check)
+            // reports what the stream really carries instead of silently claiming full chroma.
+            // Only the effective value is cleared: `chroma_444_requested` keeps the negotiation, so
+            // a later re-init on an RGB capture recovers 4:4:4 instead of being stuck at 4:2:0.
            if self.chroma_444
                && !matches!(
                    self.buffer_fmt,
@@ -1181,9 +1440,23 @@ impl Encoder for NvencD3d11Encoder {
        // despite this comment previously claiming otherwise. Since this backend encodes the
        // capturer's textures in place, exceeding the capturer's declared `pipeline_depth` lets it
        // rotate a texture out from under a live encode — torn frames, silently.
+        // FAIL SAFE when nobody told us. `set_input_ring_depth` is a defaulted trait method, so a
+        // caller that forgets it fails SILENTLY — and one did: the GameStream loop opens an encoder
+        // (`gamestream/stream.rs`) and never calls it, while the Windows IDD-push capturer declares a
+        // ring of 2 and `async_inflight_cap()` defaults to 4. That let a Moonlight session pipeline
+        // four encodes against a two-texture ring, i.e. exactly the in-place overwrite this cap
+        // exists to prevent, as torn/mixed frames and never an error.
+        //
+        // Guarding at the single point of CONSUMPTION rather than at each call site is deliberate:
+        // it covers every present and future caller, including ones that have not been written yet,
+        // whereas plumbing the setter into N loops only fixes the N sites someone remembered. An
+        // unknown ring is treated as the shallowest one any capturer in this tree declares, so the
+        // unconfigured path degrades to less pipelining — a latency cost, not corruption. Callers
+        // that DO configure it are unaffected.
+        const UNCONFIGURED_RING_DEPTH: usize = 2;
        let cap = match self.input_ring_depth {
            Some(d) => async_inflight_cap().min(d.max(1)),
-            None => async_inflight_cap(),
+            None => async_inflight_cap().min(UNCONFIGURED_RING_DEPTH),
        };
        while self.async_rt.is_some() && self.pending.len() >= cap {
            let done = {
@@ -1380,6 +1653,8 @@ impl Encoder for NvencD3d11Encoder {
        // RFI is probed once at open (`rfi_supported`); HDR SEI rides keyframes whenever the
        // session is in HDR mode. Both are the real capabilities the session glue routes on.
        EncoderCaps {
+            // The Windows capture path composites the pointer; this backend never reads `frame.cursor`.
+            blends_cursor: false,
            supports_rfi: self.rfi_supported,
            // In-band mastering/CLL is attached as keyframe SEI on HEVC/H.264 only — AV1 carries
            // it in METADATA OBUs (`HDR_MDCV`/`HDR_CLL`), which this backend doesn't emit yet
@@ -1404,61 +1679,56 @@ impl Encoder for NvencD3d11Encoder {
    }

    fn invalidate_ref_frames(&mut self, first: i64, last: i64) -> bool {
-        // No live session, the GPU can't invalidate, or a nonsense range → caller forces a full IDR.
-        // (NVENC handles are single-threaded; this runs on the encode thread, like submit/poll.)
-        if self.encoder.is_null() || !self.rfi_supported || first < 0 || first > last {
+        // No live session or the GPU can't invalidate → caller forces a full IDR. (NVENC handles
+        // are single-threaded; this runs on the encode thread, like submit/poll.) Everything else
+        // — range validity, covering-range dedup, the DPB window, the clamp — is
+        // `nvenc_core::plan_range_recovery`, one policy for both direct-NVENC backends.
+        if self.encoder.is_null() || !self.rfi_supported {
            return false;
        }
-        // Already invalidated a covering range for this loss event — no new driver calls needed,
-        // no IDR. RE-ARM the anchor though: the client re-asking means the previous recovery
-        // anchor AU may itself have been lost, and the next frame is just as clean a re-anchor
-        // (it too references only valid frames).
-        if let Some((pf, pl)) = self.last_rfi_range {
-            if first >= pf && last <= pl {
+        match plan_range_recovery(first, last, self.frame_idx, self.last_rfi_range) {
+            // Already invalidated a covering range for this loss event — no new driver calls
+            // needed, no IDR. RE-ARM the anchor though: the client re-asking means the previous
+            // recovery anchor AU may itself have been lost, and the next frame is just as clean a
+            // re-anchor (it too references only valid frames).
+            RangePlan::Covered => {
                self.pending_anchor = true;
-                return true;
+                true
            }
-        }
-        // `frame_idx` is the NEXT timestamp to assign, so the last encoded frame is `frame_idx - 1`
-        // and the DPB holds `[frame_idx - RFI_DPB, frame_idx - 1]`. A lost frame older than that
-        // can't be invalidated, so the only correct recovery is an IDR.
-        let oldest_in_dpb = self.frame_idx - RFI_DPB as i64;
-        if first < oldest_in_dpb {
-            return false;
-        }
-        // Clamp to frames we've actually encoded (don't invalidate a timestamp we never assigned).
-        let last = last.min(self.frame_idx - 1);
-        if first > last {
-            return false;
-        }
-        // Each input's `inputTimeStamp` is `frame_idx`, which `submit_indexed` pins to the WIRE
-        // frame index the AU carries — so the client's lost-frame range maps 1:1 onto the
-        // timestamps NVENC invalidates here, and stays 1:1 across encoder rebuilds/resets (an
-        // internal counter would desync on the first adaptive-bitrate rebuild and RFI would then
-        // clamp every range into first > last, silently degrading to IDR-only forever).
-        // SAFETY: `invalidate_ref_frames` is a function pointer from the runtime-loaded `EncodeApi` table.
-        // `self.encoder` was checked non-null at the top of this fn and is the live session; this runs
-        // on the encode thread (like submit/poll), so there is no concurrent NVENC use. Each `ts` was
-        // clamped to `[oldest_in_dpb, frame_idx - 1]` above, so it names a frame still in the session's
-        // DPB; the call passes only that `u64` timestamp (no struct), so there is no struct-size or
-        // lifetime concern.
-        unsafe {
-            for ts in first..=last {
-                if (api().invalidate_ref_frames)(self.encoder, ts as u64)
-                    .nv_ok()
-                    .is_err()
-                {
-                    return false; // any failure → fall back to IDR
+            RangePlan::Decline => false,
+            RangePlan::Invalidate { first, last } => {
+                // Each input's `inputTimeStamp` is `frame_idx`, which `submit_indexed` pins to the
+                // WIRE frame index the AU carries — so the client's lost-frame range maps 1:1 onto
+                // the timestamps NVENC invalidates here, and stays 1:1 across encoder
+                // rebuilds/resets (an internal counter would desync on the first adaptive-bitrate
+                // rebuild and RFI would then clamp every range into first > last, silently
+                // degrading to IDR-only forever).
+                // SAFETY: `invalidate_ref_frames` is a function pointer from the runtime-loaded
+                // `EncodeApi` table. `self.encoder` was checked non-null at the top of this fn and
+                // is the live session; this runs on the encode thread (like submit/poll), so there
+                // is no concurrent NVENC use. The plan clamped each `ts` to
+                // `[oldest_in_dpb, frame_idx - 1]`, so it names a frame still in the session's
+                // DPB; the call passes only that `u64` timestamp (no struct), so there is no
+                // struct-size or lifetime concern.
+                unsafe {
+                    for ts in first..=last {
+                        if (api().invalidate_ref_frames)(self.encoder, ts as u64)
+                            .nv_ok()
+                            .is_err()
+                        {
+                            return false; // any failure → fall back to IDR
+                        }
+                    }
                }
+                self.last_rfi_range = Some((first, last));
+                // The next submitted frame is the first one encoded after the invalidation — the
+                // clean re-anchor P-frame. Arm the tag so its AU ships with `recovery_anchor` and
+                // the client lifts its post-loss freeze on it (instead of waiting ~1 s for the
+                // cooldown-suppressed IDR fallback).
+                self.pending_anchor = true;
+                true
            }
        }
-        self.last_rfi_range = Some((first, last));
-        // The next submitted frame is the first one encoded after the invalidation — the clean
-        // re-anchor P-frame. Arm the tag so its AU ships with `recovery_anchor` and the client
-        // lifts its post-loss freeze on it (instead of waiting ~1 s for the cooldown-suppressed
-        // IDR fallback).
-        self.pending_anchor = true;
-        true
    }

    fn poll(&mut self) -> Result<Option<EncodedFrame>> {
@@ -1552,6 +1822,15 @@ impl Encoder for NvencD3d11Encoder {
            self.bitrate_bps = bps;
            return true;
        }
+        // Cached codec-level ceiling: clamp the target BEFORE the driver call, so a known
+        // overshoot retargets to the ceiling IN PLACE instead of bouncing off the driver into
+        // the caller's full-rebuild fallback (an IDR plus ~half a second of session churn per
+        // ABR overshoot on the pre-cache path). The caller reads the clamp back through
+        // [`Encoder::applied_bitrate_bps`].
+        let bps = match cached_ceiling(&self.ceiling_key(self.split_mode)) {
+            Some(ceiling) => bps.min(ceiling),
+            None => bps,
+        };
        // SAFETY: `inited` ⟹ `self.encoder` is the live session and this runs on the encode
        // thread between submit/poll (`nvEncReconfigureEncoder` is a submit-side call, the
        // sanctioned side of the two-thread async split — the retrieve thread only ever locks
@@ -1600,6 +1879,12 @@ impl Encoder for NvencD3d11Encoder {
        }
    }

+    fn applied_bitrate_bps(&self) -> Option<u64> {
+        // `bitrate_bps` is the post-clamp truth: the open path's ceiling search and the
+        // reconfigure path's cache clamp both write what the session ACTUALLY targets.
+        Some(self.bitrate_bps)
+    }
+
    fn flush(&mut self) -> Result<()> {
        Ok(()) // P1/ULL + frameIntervalP=1: each submit yields its AU; no internal queue to drain.
    }
@@ -1645,6 +1930,11 @@ pub fn probe_can_encode_10bit(codec: Codec) -> bool {
 /// on any failure (no loadable NVENC, no device, failed open) — the honest answer for a
 /// capability that couldn't be confirmed.
 fn probe_encode_cap(codec: Codec, cap: nv::NV_ENC_CAPS) -> bool {
+    // Same exclusion as `init_session`: this opens a real (throwaway) session, so it must never
+    // overlap a zombie reap that could be destroying the very address the driver hands us.
+    let _gate = DRIVER_SESSION_GATE
+        .lock()
+        .unwrap_or_else(|p| p.into_inner());
    use windows::Win32::Foundation::HMODULE;
    use windows::Win32::Graphics::Direct3D::{
        D3D_DRIVER_TYPE_HARDWARE, D3D_DRIVER_TYPE_UNKNOWN, D3D_FEATURE_LEVEL_11_0,
@@ -1723,6 +2013,9 @@ fn probe_encode_cap(codec: Codec, cap: nv::NV_ENC_CAPS) -> bool {
            }
            return false;
        }
+        // Availability probe, but a real session open all the same: it proves the driver accepted
+        // this build's version word, which is what rules a skew out later (see `nvenc_status`).
+        nvenc_status::note_session_opened();
        let mut param = nv::NV_ENC_CAPS_PARAM {
            version: nv::NV_ENC_CAPS_PARAM_VER,
            capsToQuery: cap,
@@ -189,12 +189,15 @@ impl PyroWaveEncoder {
        if !chroma444 && (width % 2 != 0 || height % 2 != 0) {
            bail!("pyrowave 4:2:0 needs even dimensions (got {width}x{height})");
        }
-        if chroma444 && !crate::pyrowave_mode_fits_rdo(width, height, true) {
-            // The negotiator downgrades these modes to 4:2:0 pre-Welcome; refuse if one
-            // slips through (e.g. the lab override) rather than wrap the RDO block index.
+        // Checked against the chroma actually being opened, NOT hardcoded 4:4:4 — see the Linux
+        // twin (`enc/linux/pyrowave.rs`) for the full rationale: the negotiator's 4:4:4 → 4:2:0
+        // downgrade hands oversized modes to this open AS 4:2:0, so the old `chroma444`-gated
+        // check was skipped exactly when it was needed.
+        if !crate::pyrowave_mode_fits_rdo(width, height, chroma444) {
            bail!(
-                "pyrowave 4:4:4 at {width}x{height} exceeds the rate controller's 16-bit \
-                 block index (see pyrowave-sys patches/0002 note) — use 4:2:0 at this size"
+                "pyrowave {} at {width}x{height} exceeds the rate controller's 16-bit block \
+                 index (see pyrowave-sys patches/0002 note) — lower the resolution",
+                if chroma444 { "4:4:4" } else { "4:2:0" }
            );
        }
        let fps = fps.max(1);
@@ -294,8 +297,10 @@ impl PyroWaveEncoder {
    /// Import one capturer plane D3D11 texture (`R8_UNORM` Y or `R8G8_UNORM` CbCr) into pyrowave's
    /// Vulkan device. Creates a fresh shared NT handle from the texture (the capturer marked the ring
    /// `SHARED | SHARED_NTHANDLE`); `pyrowave_image_create` takes ownership of the handle and closes
-    /// it on import. Single/two-component textures import reliably on NVIDIA at any size — unlike a
-    /// planar NV12 — so no MUTABLE_FORMAT / planar-layout workaround is involved.
+    /// it on SUCCESSFUL import only (pyrowave-sys patch 0006 — same contract as the fence import in
+    /// [`Self::import_fence`]), so this fn closes it on every failure return. Single/two-component
+    /// textures import reliably on NVIDIA at any size — unlike a planar NV12 — so no
+    /// MUTABLE_FORMAT / planar-layout workaround is involved.
    ///
    /// # Safety
    /// `texture` must be a live `ID3D11Texture2D` of format `vk_format`, sized `w`×`h`, created
@@ -346,7 +351,11 @@ impl PyroWaveEncoder {
        };
        let mut image: pw::pyrowave_image = std::ptr::null_mut();
        if let Err(e) = pw_check(pw::pyrowave_image_create(&info, &mut image), "image_create") {
-            // pyrowave only closes the handle on a SUCCESSFUL import — close it ourselves on failure.
+            // pyrowave consumes the handle ONLY on a successful import (pyrowave-sys patch 0006
+            // pinned this at the API's success boundary) — so on EVERY failure return the handle
+            // is still ours and this close is the single one. Before the patch, an
+            // allocate-stage failure inside Granite had already closed it and this was a double
+            // close of a possibly-recycled handle value.
            let _ = CloseHandle(handle);
            return Err(e);
        }
@@ -416,7 +425,9 @@ impl PyroWaveEncoder {
            pw::pyrowave_sync_object_create(&info, &mut sync),
            "sync_object_create",
        ) {
-            // pyrowave only closes the handle on a SUCCESSFUL import — close the dup on failure.
+            // pyrowave only closes the handle on a SUCCESSFUL import (Granite's semaphore import
+            // has always had these semantics; patch 0006 made the image import match) — close
+            // the dup on failure.
            let _ = CloseHandle(dup);
            return Err(e);
        }
@@ -678,6 +689,8 @@ impl Encoder for PyroWaveEncoder {
        // after the caps() default was written — a hardcoded `default()` here mis-reports a 4:4:4
        // open as 4:2:0 and fires a spurious "chroma disagrees with the negotiated Welcome" warn).
        EncoderCaps {
+            // The Windows capturer composites the pointer itself; this backend never reads it.
+            blends_cursor: false,
            chroma_444: self.chroma444,
            ..EncoderCaps::default()
        }
@@ -519,9 +519,20 @@ fn build_params(cfg: &EncodeConfig) -> ParamSet {
        }
        b.WhitePointX = m.white_point[0];
        b.WhitePointY = m.white_point[1];
-        // Units diverge on the max: VPL wants whole cd/m² (HdrMeta carries 0.0001 cd/m²); the
-        // min is 0.0001 cd/m² on both sides.
-        b.MaxDisplayMasteringLuminance = m.max_display_mastering_luminance / 10_000;
+        // BOTH luminance fields are 0.0001 cd/m² here — do NOT scale the max.
+        //
+        // The `/10_000` this used to carry followed the VPL header's *video-processing* unit
+        // (whole cd/m², which is right for VPP), but on the ENCODE path the runtime passes these
+        // straight into the ITU-T H.265 Annex D mastering-display SEI, whose unit is 0.0001 cd/m².
+        // Dividing therefore under-reported peak brightness by 10,000x. Confirmed in a real
+        // bitstream on Intel UHD 750 (2026-07-25): SEI 137 carried
+        // max_display_mastering_luminance = 1000, i.e. **0.1 nits** for a 1000-nit display, while
+        // the undivided min = 500 (0.05 nits) was correct — the asymmetry was the tell. A client
+        // tone-mapping against 0.1 nits crushes the image.
+        //
+        // Unscaled also matches every sibling: `MinDisplayMasteringLuminance` right below,
+        // `pf_frame::hdr::hevc_mastering_display_sei`, and the AMF backend.
+        b.MaxDisplayMasteringLuminance = m.max_display_mastering_luminance;
        b.MinDisplayMasteringLuminance = m.min_display_mastering_luminance;
        b
    });
@@ -690,13 +701,17 @@ impl Inner {
    }

    fn take_bs(&mut self) -> Box<BsBuf> {
-        match self.bs_pool.pop() {
-            Some(mut b) => {
+        // Pooled buffers were sized by whatever `bs_bytes` was when they were allocated. A bitrate
+        // retarget raises the driver's worst-case AU size, so a recycled buffer can be SMALLER than
+        // the current requirement — hand those back to the allocator instead of letting the runtime
+        // write an AU into a short buffer.
+        while let Some(mut b) = self.bs_pool.pop() {
+            if b.mfx.MaxLength as usize >= self.bs_bytes {
                b.recycle();
-                b
+                return b;
            }
-            None => BsBuf::new(self.bs_bytes),
        }
+        BsBuf::new(self.bs_bytes)
    }
 }

@@ -774,7 +789,12 @@ impl QsvEncoder {
        if codec == Codec::Av1 && !probe_can_encode(Codec::Av1) {
            bail!("this GPU/driver declined AV1 encode (DG2/Arc or MTL+ required) — QSV probe");
        }
-        let ten_bit = bit_depth >= 10 || matches!(format, PixelFormat::P010 | PixelFormat::Rgb10a2);
+        // Depth follows the delivered pixels, not the negotiated depth ([`crate::ten_bit_input`]).
+        // With the old `bit_depth >= 10 || …` shape a 10-bit-negotiated session over an 8-bit
+        // capture derived `expected = P010` and hit the bail below, ending the session at open —
+        // and taking the ffmpeg fallback with it, since that path had the same defect in a worse
+        // form (it accepted the open and then failed every frame).
+        let ten_bit = crate::ten_bit_input(format, bit_depth);
        if ten_bit && codec == Codec::H264 {
            bail!("native QSV: 10-bit is HEVC/AV1-only (H.264 High10 is not negotiated)");
        }
@@ -878,7 +898,8 @@ impl QsvEncoder {
        if sts > vpl::MFX_ERR_NONE {
            tracing::debug!(status = sts_name(sts), "QSV Init returned a warning");
        }
-        let ir_active = cfg.intra_refresh && set.co2.is_some();
+        // Whether we ASKED for intra-refresh. Not the same as getting it — confirmed below.
+        let ir_requested = cfg.intra_refresh && set.co2.is_some();
        // The driver's own answer for the worst-case AU size.
        // SAFETY: `session` is live; `got` and its (empty) ext chain outlive the call.
        let bs_bytes = unsafe {
@@ -892,6 +913,57 @@ impl QsvEncoder {
            let kb = enc_of(m).BufferSizeInKB as usize;
            (kb * mult * 1000).max(256 * 1024)
        };
+        // Intra-refresh HONESTY: ask the driver what it actually installed instead of trusting that
+        // it took what we sent. Both `Query` and `Init` can return MFX_WRN_INCOMPATIBLE_VIDEO_PARAM
+        // — a WARNING, which the code above accepts — while silently dropping the wave. Confirmed
+        // on Intel UHD 750 (2026-07-25): H.264 reports exactly that, `GetVideoParam` comes back with
+        // `IntRefType = 0`, and `ir_active` still claimed true. H.265 on the same GPU is genuinely
+        // active, so this is per-codec and cannot be decided statically.
+        //
+        // That lie is not cosmetic: `ir_active` feeds `EncoderCaps::intra_refresh`, so the session
+        // advertises gradual refresh to the client and then never emits it — the client stops
+        // requesting the IDRs it would otherwise have asked for on loss, and a lost frame is
+        // concealed instead of repaired.
+        //
+        // Deliberately a SEPARATE, best-effort query rather than a buffer chained onto the
+        // `BufferSizeInKB` call above (which is what the audit finding suggested): that value is
+        // load-bearing for every bitstream allocation, and a runtime that dislikes the chained
+        // buffer would take it down with the readback. A failure here costs only the verdict, and
+        // is resolved conservatively.
+        let ir_active = if ir_requested {
+            // SAFETY: `session` is live on this thread; `got` and `co2_out` (with `ptrs` holding the
+            // only reference to it) all outlive the synchronous call, and the runtime writes back
+            // only into the buffer whose header we stamped.
+            let confirmed = unsafe {
+                let mut got: vpl::mfxVideoParam = std::mem::zeroed();
+                let mut co2_out: vpl::mfxExtCodingOption2 = std::mem::zeroed();
+                co2_out.Header.BufferId = vpl::MFX_EXTBUFF_CODING_OPTION2 as u32;
+                co2_out.Header.BufferSz = std::mem::size_of::<vpl::mfxExtCodingOption2>() as u32;
+                let mut ptrs: [*mut vpl::mfxExtBuffer; 1] = [&mut co2_out.Header as *mut _];
+                got.ExtParam = ptrs.as_mut_ptr();
+                got.NumExtParam = 1;
+                let sts = vpl::MFXVideoENCODE_GetVideoParam(session, &mut got);
+                if sts < vpl::MFX_ERR_NONE {
+                    tracing::debug!(
+                        status = sts_name(sts),
+                        "QSV: could not read back CodingOption2 — trusting the intra-refresh request"
+                    );
+                    true
+                } else {
+                    co2_out.IntRefType != 0
+                }
+            };
+            if !confirmed {
+                tracing::warn!(
+                    codec = ?cfg.codec,
+                    "QSV silently dropped intra-refresh (GetVideoParam reports IntRefType=0) — \
+                     advertising it OFF so the client keeps asking for IDRs on loss"
+                );
+            }
+            confirmed
+        } else {
+            false
+        };
        Ok((ltr_active, ir_active, bs_bytes))
    }

@@ -1344,36 +1416,30 @@ impl Encoder for QsvEncoder {
        if !self.ltr_active || first < 0 || first > last {
            return false;
        }
-        // Taint sweep BEFORE picking the anchor: an LTR marked at-or-after the loss start was
-        // encoded inside the client's corrupt window — the client either never received it or
-        // decoded it against a broken reference chain. Serving it as "known-good" on a LATER
-        // loss ships corruption as the recovery anchor, and every subsequent mark re-samples
-        // the soup — the sustained-loss field failure where the picture never healed. Dropped
-        // slots stay dropped; the cadence re-marks a clean frame within ~1/4 s.
-        //
-        // Mark tainted rather than clearing: `ltr_slots` mirrors the HARDWARE DPB, and nulling an
-        // entry issues no VPL call — the frame stays marked long-term in the encoder. Clearing it
-        // made the rejection list below (which iterates the post-sweep mirror and only names `Some`
-        // slots) silently SKIP the one entry the sweep exists to distrust, so the recovery frame
-        // could still predict from it. With two slots the "exactly one swept" case is the modal
-        // one, and it was the broken one.
-        for (slot, marked) in self.ltr_slots.iter().enumerate() {
-            if marked.is_some_and(|idx| idx >= first) {
-                self.ltr_tainted[slot] = true;
+        // The taint-sweep + anchor-pick POLICY lives in `rfi::plan_slot_recovery` (one decision
+        // shared with AMF and Vulkan Video). This backend's mechanism: distrust is a SEPARATE
+        // `ltr_tainted` flag, never a cleared mirror slot — `ltr_slots` mirrors the HARDWARE DPB,
+        // and nulling an entry issues no VPL call, so the frame stays marked long-term in the
+        // encoder. Clearing it made the rejection list below (which iterates the mirror and only
+        // names `Some` slots) silently SKIP the one entry the sweep exists to distrust, so the
+        // recovery frame could still predict from it. With two slots the "exactly one swept" case
+        // is the modal one, and it was the broken one. The `!ltr_tainted` view filter below is
+        // what persists that distrust across loss events (this call's taints are excluded from
+        // this call's anchor by the policy itself).
+        let view: Vec<(usize, i64)> = self
+            .ltr_slots
+            .iter()
+            .enumerate()
+            .filter(|&(slot, _)| !self.ltr_tainted[slot])
+            .filter_map(|(s, m)| m.map(|w| (s, w)))
+            .collect();
+        let plan = super::rfi::plan_slot_recovery(&view, first);
+        for (slot, tainted) in self.ltr_tainted.iter_mut().enumerate() {
+            if plan.tainted & (1 << slot) != 0 {
+                *tainted = true;
            }
        }
-        let mut best: Option<(usize, i64)> = None;
-        for (slot, marked) in self.ltr_slots.iter().enumerate() {
-            if self.ltr_tainted[slot] {
-                continue; // still in the DPB, but encoded inside the corrupt window
-            }
-            if let Some(idx) = *marked {
-                if idx < first && best.is_none_or(|(_, b)| idx > b) {
-                    best = Some((slot, idx));
-                }
-            }
-        }
-        match best {
+        match plan.anchor {
            Some((slot, ltr_frame)) => {
                self.pending_force = Some(slot);
                tracing::info!(
@@ -1402,6 +1468,8 @@ impl Encoder for QsvEncoder {

    fn caps(&self) -> EncoderCaps {
        EncoderCaps {
+            // As Windows NVENC: the capturer composites; this backend never reads `frame.cursor`.
+            blends_cursor: false,
            supports_rfi: self.ltr_active,
            // In-band mastering/CLL at IDR (HEVC prefix SEI / AV1 metadata OBU); AVC sessions
            // are never HDR.
@@ -1576,6 +1644,38 @@ impl Encoder for QsvEncoder {
            self.bitrate_bps = old;
            return false;
        }
+        // Re-read the driver's worst-case AU size. `Reset` re-derives `BufferSizeInKB` from the new
+        // rate, and a step-up can raise it well above what `init_encode` computed — leaving
+        // `bs_bytes` (and every buffer already in the pool) sized for the OLD, lower bitrate. This
+        // mirrors `init_encode`, which asks the same question post-Init; `take_bs` drops any pooled
+        // buffer that is now too small.
+        // SAFETY: `session` is live on this thread and drained above; `got` and its (empty) ext
+        // chain outlive the synchronous call.
+        let refreshed = unsafe {
+            let mut got: vpl::mfxVideoParam = std::mem::zeroed();
+            let sts = vpl::MFXVideoENCODE_GetVideoParam(session, &mut got);
+            (sts >= vpl::MFX_ERR_NONE).then(|| {
+                let m = &mut got.__bindgen_anon_1.mfx;
+                let mult = m.BRCParamMultiplier.max(1) as usize;
+                let kb = enc_of(m).BufferSizeInKB as usize;
+                (kb * mult * 1000).max(256 * 1024)
+            })
+        };
+        match (refreshed, self.inner.as_mut()) {
+            (Some(bytes), Some(inner)) if bytes > inner.bs_bytes => {
+                tracing::debug!(
+                    old_bytes = inner.bs_bytes,
+                    new_bytes = bytes,
+                    mbps = bps / 1_000_000,
+                    "QSV retarget raised the worst-case AU size — resizing the bitstream pool"
+                );
+                inner.bs_bytes = bytes;
+            }
+            (None, _) => tracing::warn!(
+                "QSV retarget: GetVideoParam failed, keeping the previous AU buffer size"
+            ),
+            _ => {}
+        }
        true
    }

@@ -5,9 +5,11 @@
 //! VA surface). One [`Encoder`] trait, selected in [`open_video`]. Extracted into a subsystem crate
 //! (plan §W6): depends on the shared frame vocabulary (`pf-frame`) + zero-copy plumbing
 //! (`pf-zerocopy`), never on capture — the capture→encode edge is one-way.
-// Scaffold: some backend paths + trait defaults are defined ahead of the per-feature build that
-// uses them (mirrors the host crate root's allow before the extraction).
-#![allow(dead_code)]
+// NOTE: no crate-wide `#![allow(dead_code)]`. It was inherited from the pre-extraction host crate
+// root as scaffolding for backend paths defined ahead of the build that used them, but a census
+// across every feature combination on both platforms found it was hiding exactly two items — so it
+// bought nothing and blinded the crate to future rot. Genuinely test-only helpers carry
+// `#[cfg(test)]` instead.
 // Every unsafe block in this module tree carries a `// SAFETY:` proof; enforce it (unsafe-proof
 // program). As a parent module this also covers the child modules (windows/linux backends).
 #![deny(clippy::undocumented_unsafe_blocks)]
@@ -39,11 +41,14 @@ impl Codec {
        // per-session instead (the host `session_plan::SessionPlan::output_format`) — the EGL→CUDA
        // frames the `auto` GPU path would deliver are NVENC-only. Only a software/GPU-less pref
        // keeps the bit off (no Vulkan device to open).
+        // Resolved ONCE for every gate below (pyro bit, software pin, vulkan ceiling, the
+        // zero-copy plane): this fn backs the codec advertisement on polled paths, and the
+        // resolver's auto arm samples live GPU-preference state — four samples per call was the
+        // probe-amplification the WP7.6 diff critic caught.
+        #[cfg(target_os = "linux")]
+        let backend = linux_resolved_backend();
        #[cfg(all(target_os = "linux", feature = "pyrowave"))]
-        let pyro = if !matches!(
-            pf_host_config::config().encoder_pref.as_str(),
-            "software" | "sw" | "openh264"
-        ) {
+        let pyro = if backend != LinuxBackend::Software {
            punktfunk_core::quic::CODEC_PYROWAVE
        } else {
            0u8
@@ -70,19 +75,41 @@ impl Codec {
                | punktfunk_core::quic::CODEC_AV1;
            #[cfg(target_os = "linux")]
            {
-                if matches!(
-                    pf_host_config::config().encoder_pref.as_str(),
-                    "software" | "sw" | "openh264"
-                ) {
+                if backend == LinuxBackend::Software {
                    return punktfunk_core::quic::CODEC_H264;
                }
-                if linux_zero_copy_is_vaapi() {
+                // A pref that FORCES the raw Vulkan Video backend can only ever serve what that
+                // backend encodes: `open_video_backend`'s `vulkan` arm bails outright for anything
+                // that is not HEVC/AV1 ("the Vulkan Video encoder supports HEVC + AV1; the session
+                // negotiated {codec:?}"). Advertising H.264 there let a client negotiate it and die
+                // at encoder open. Without the `vulkan-encode` feature that arm cannot open at all,
+                // so the pref advertises nothing.
+                //
+                // This is a CEILING intersected with the device probe below, never a replacement
+                // for it: pinning a static HEVC|AV1 would ADD AV1 on the AMD/Intel hosts whose probe
+                // currently withholds it (pre-RDNA3, pre-Arc), i.e. it would re-create this very bug
+                // for a different codec. Intersecting can only narrow.
+                let pref_ceiling: u8 = match backend {
+                    // Feature-blindness rider (WP7.6 critics): the resolver knows the PREF is
+                    // vulkan; only this cfg! knows the BUILD can open it. Dropping the inner
+                    // branch would advertise HEVC|AV1 on a featureless build — the
+                    // advertise-then-die-at-open bug this ceiling exists to prevent.
+                    LinuxBackend::Vulkan => {
+                        if cfg!(feature = "vulkan-encode") {
+                            punktfunk_core::quic::CODEC_HEVC | punktfunk_core::quic::CODEC_AV1
+                        } else {
+                            0
+                        }
+                    }
+                    _ => GPU_SUPERSET,
+                };
+                if linux_zero_copy_is_vaapi_for(backend) {
                    if let Some(m) = vaapi_codec_support().wire_mask() {
-                        return m;
+                        return m & pref_ceiling;
                    }
                }
                // NVENC (static superset, like GameStream) — or an empty VAAPI probe (see above).
-                GPU_SUPERSET
+                GPU_SUPERSET & pref_ceiling
            }
            #[cfg(target_os = "windows")]
            {
@@ -175,6 +202,23 @@ pub fn open_video(
            },
        }
    };
+    // The session asked for a composited pointer; say so loudly if the backend that actually opened
+    // cannot deliver one. `cursor_blend` was a REQUEST with no answer for most of this crate's life
+    // (`let _ = cursor_blend;` below), and the result was a stream with no mouse cursor and nothing
+    // in the logs — confirmed on the VAAPI dmabuf path and the libav-NVENC CUDA path.
+    //
+    // A warning is deliberately all this does. `open_video` cannot re-plan capture, so refusing here
+    // would only trade a missing pointer for a dead session; the host owns `plan.cursor_blend` and is
+    // the only layer that can fall back to capturer-side compositing. This makes the condition
+    // visible and queryable (`EncoderCaps::blends_cursor`) so that decision can be made upstream.
+    if cursor_blend && !inner.caps().blends_cursor {
+        tracing::warn!(
+            backend,
+            "session negotiated a composited cursor but this encode backend does not blend \
+             CapturedFrame::cursor — the pointer will be MISSING from the stream unless the \
+             capturer composites it"
+        );
+    }
    Ok(Box::new(TrackedEncoder {
        inner,
        _session: pf_gpu::session_begin(gpu),
@@ -241,6 +285,12 @@ impl Encoder for TrackedEncoder {
    fn reconfigure_bitrate(&mut self, bps: u64) -> bool {
        self.inner.reconfigure_bitrate(bps)
    }
+    // Forwarded (the same trap class as `set_wire_chunking`): the unforwarded default `None`
+    // would tell the session loop "no encoder truth here" and it would keep pacing/acking the
+    // requested rate even where NVENC clamped to the codec-level ceiling.
+    fn applied_bitrate_bps(&self) -> Option<u64> {
+        self.inner.applied_bitrate_bps()
+    }
    fn flush(&mut self) -> Result<()> {
        self.inner.flush()
    }
@@ -253,6 +303,224 @@ impl Encoder for TrackedEncoder {
 #[cfg(any(target_os = "linux", target_os = "windows"))]
 const SW_BITRATE_CEIL: u64 = 100_000_000;

+/// The Linux half of [`open_video_backend`], with the pref INJECTED — the seam exists so the
+/// dispatch is testable without mutating process env (`set_var` races `getenv` in parallel
+/// tests) or fighting the once-latched `pf_host_config::config()`. Production takes the
+/// config pref via the caller; shared validation (dims/fps/chroma degrade) stays there too.
+#[cfg(target_os = "linux")]
+#[allow(clippy::too_many_arguments)]
+fn open_video_backend_linux(
+    pref: &str,
+    codec: Codec,
+    format: PixelFormat,
+    width: u32,
+    height: u32,
+    fps: u32,
+    bitrate_bps: u64,
+    cuda: bool,
+    bit_depth: u8,
+    chroma: ChromaFormat,
+    cursor_blend: bool,
+) -> Result<(Box<dyn Encoder>, &'static str)> {
+    // A NEGOTIATED PyroWave session (client advertised + preferred it, plan §3) routes
+    // straight to that backend — the PUNKTFUNK_ENCODER pref below stays a lab override.
+    if codec == Codec::PyroWave {
+        #[cfg(feature = "pyrowave")]
+        {
+            return pyrowave::PyroWaveEncoder::open(width, height, fps, bitrate_bps, chroma)
+                .map(|e| (Box::new(e) as Box<dyn Encoder>, "pyrowave"));
+        }
+        #[cfg(not(feature = "pyrowave"))]
+        anyhow::bail!(
+            "session negotiated PyroWave but this host was built without --features \
+             punktfunk-host/pyrowave (the advertisement bit should not have been set)"
+        );
+    }
+    // Pick the GPU encode backend. NVIDIA → NVENC/CUDA (the original path, unchanged);
+    // AMD/Intel → VAAPI (one libavcodec backend for both). Auto-detect by default so a single
+    // Linux binary serves any GPU; `PUNKTFUNK_ENCODER` forces a specific backend (and surfaces
+    // its errors crisply instead of silently trying the other).
+    // AMD/Intel opener. Default = libav VAAPI. With `--features vulkan-encode` +
+    // PUNKTFUNK_VULKAN_ENCODE, an HEVC session instead opens the raw Vulkan Video backend (real
+    // RFI loss recovery the VAAPI path can't express); a failed open falls back to VAAPI so the
+    // stream never dies over the new path. `format`/`bit_depth`/`chroma` only matter to VAAPI —
+    // the Vulkan backend imports the dmabuf and does its own 8-bit 4:2:0 CSC.
+    let open_amd_intel = || -> Result<(Box<dyn Encoder>, &'static str)> {
+        // An HDR session (10-bit + a PQ/BT.2020 capture format) must skip the Vulkan Video
+        // backend — it hardcodes an 8-bit 4:2:0 BT.709 CSC — and take the libav VAAPI path,
+        // which has the P010/Main10/PQ wiring. SDR sessions keep the Vulkan default.
+        #[cfg(feature = "vulkan-encode")]
+        if matches!(codec, Codec::H265 | Codec::Av1)
+            && vulkan_encode_enabled()
+            && !(bit_depth == 10 && format.is_hdr_rgb10())
+        {
+            match vulkan_video::VulkanVideoEncoder::open(
+                codec,
+                format,
+                width,
+                height,
+                fps,
+                bitrate_bps,
+                cursor_blend,
+            ) {
+                Ok(e) => {
+                    tracing::info!(
+                        codec = ?codec,
+                        "Linux Vulkan Video encode (real RFI via DPB reference slots) — \
+                         set PUNKTFUNK_VULKAN_ENCODE=0 for libav VAAPI"
+                    );
+                    return Ok((Box::new(e) as Box<dyn Encoder>, "vulkan"));
+                }
+                // Native NV12 (PUNKTFUNK_PIPEWIRE_NV12 capture) has no VAAPI fallback:
+                // libav's dmabuf lane would import the two-plane buffer as packed RGB
+                // (silent garbage) and its CPU lane bails per frame — die crisply instead.
+                Err(e) if format == PixelFormat::Nv12 => {
+                    return Err(e.context(
+                        "Vulkan Video open failed on a native-NV12 capture \
+                         — no VAAPI fallback exists; set PUNKTFUNK_PIPEWIRE_NV12=0 to \
+                         restore the packed-RGB negotiation",
+                    ));
+                }
+                Err(e) => tracing::warn!(
+                    error = %format!("{e:#}"),
+                    "Vulkan Video encode open failed — falling back to libav VAAPI"
+                ),
+            }
+        }
+        // Same rule when the Vulkan backend was never eligible (H264 session,
+        // PUNKTFUNK_VULKAN_ENCODE=0, or a build without the feature).
+        if format == PixelFormat::Nv12 {
+            anyhow::bail!(
+                "native NV12 capture requires the Vulkan Video encoder (HEVC/AV1 \
+                 session, --features vulkan-encode, PUNKTFUNK_VULKAN_ENCODE not 0) — this \
+                 session resolved to libav VAAPI; set PUNKTFUNK_PIPEWIRE_NV12=0 to restore \
+                 the packed-RGB negotiation"
+            );
+        }
+        vaapi::VaapiEncoder::open(
+            codec,
+            format,
+            width,
+            height,
+            fps,
+            bitrate_bps,
+            bit_depth,
+            chroma,
+        )
+        .map(|e| (Box::new(e) as Box<dyn Encoder>, "vaapi"))
+    };
+    let open_nvidia = || -> Result<(Box<dyn Encoder>, &'static str)> {
+        open_nvenc_probed(
+            codec,
+            format,
+            width,
+            height,
+            fps,
+            bitrate_bps,
+            cuda,
+            bit_depth,
+            chroma,
+            cursor_blend,
+        )
+        .map(|e| (e, "nvenc"))
+    };
+    // WP7.6: the dispatch consumes the SAME resolver the capability mirrors consult (the
+    // Windows shape — `windows_resolved_backend` + `match backend` below) so the alias table
+    // exists once. Arm bodies and their open-site labels are untouched.
+    match resolve_linux_backend(pref, linux_auto_is_vaapi, cuda) {
+        Some(LinuxBackend::Nvenc) => open_nvidia(),
+        Some(LinuxBackend::AmdIntel) => open_amd_intel(),
+        // Force the raw Vulkan Video HEVC backend (real RFI). Needs `--features vulkan-encode`.
+        Some(LinuxBackend::Vulkan) => {
+            #[cfg(feature = "vulkan-encode")]
+            {
+                if !matches!(codec, Codec::H265 | Codec::Av1) {
+                    anyhow::bail!(
+                        "the Vulkan Video encoder supports HEVC + AV1; the session negotiated {codec:?}"
+                    );
+                }
+                vulkan_video::VulkanVideoEncoder::open(
+                    codec,
+                    format,
+                    width,
+                    height,
+                    fps,
+                    bitrate_bps,
+                    cursor_blend,
+                )
+                .map(|e| (Box::new(e) as Box<dyn Encoder>, "vulkan"))
+            }
+            #[cfg(not(feature = "vulkan-encode"))]
+            {
+                let _ = (format, bit_depth, chroma);
+                anyhow::bail!(
+                    "PUNKTFUNK_ENCODER=vulkan requires a build with --features vulkan-encode"
+                )
+            }
+        }
+        // PyroWave — the opt-in wired-LAN intra-only wavelet codec. Explicit-only, and
+        // EXPERIMENTAL until CODEC_PYROWAVE negotiation lands (plan Phase 2): no shipping
+        // client can decode the stream yet, so this arm exists for host-side bring-up and
+        // latency work only. Vendor-agnostic (any Vulkan 1.3 GPU); ignores the negotiated
+        // codec — every AU is an independently-decodable wavelet frame.
+        Some(LinuxBackend::Pyrowave) => {
+            #[cfg(feature = "pyrowave")]
+            {
+                tracing::warn!(
+                    ?codec,
+                    "PUNKTFUNK_ENCODER=pyrowave forces the all-intra wavelet stream \
+                     regardless of the negotiated codec — only a pyrowave-feature client \
+                     that ALSO preferred CODEC_PYROWAVE can display it (lab override; \
+                     normal sessions negotiate it instead)"
+                );
+                // The lab override forces the wavelet stream onto a session negotiated for
+                // another codec — that session's chroma may be HEVC-4:4:4, which the
+                // pyrowave encoder doesn't do yet, so pin the override to 4:2:0.
+                pyrowave::PyroWaveEncoder::open(
+                    width,
+                    height,
+                    fps,
+                    bitrate_bps,
+                    ChromaFormat::Yuv420,
+                )
+                .map(|e| (Box::new(e) as Box<dyn Encoder>, "pyrowave"))
+            }
+            #[cfg(not(feature = "pyrowave"))]
+            {
+                anyhow::bail!(
+                    "PUNKTFUNK_ENCODER=pyrowave requires a build with --features punktfunk-host/pyrowave"
+                )
+            }
+        }
+        // GPU-less software H.264 (openh264) — for a headless / GPU-lost box. Explicit-only:
+        // `auto` never picks it (a box with `/dev/nvidiactl` present but a dead driver would
+        // otherwise wrongly resolve to NVENC). Needs H.264 (openh264 emits only that) and a CPU
+        // RGB frame, which the capturer delivers because the software backend resolves `gpu=false`.
+        Some(LinuxBackend::Software) => {
+            if codec != Codec::H264 {
+                anyhow::bail!(
+                    "the software encoder emits H.264 only; the session negotiated {codec:?} \
+                     (a client must advertise CODEC_H264 to reach a software host)"
+                );
+            }
+            let _ = (cuda, bit_depth); // software path is CPU + 8-bit only
+            sw::OpenH264Encoder::open(
+                format,
+                width,
+                height,
+                fps,
+                bitrate_bps.min(SW_BITRATE_CEIL),
+            )
+            .map(|e| (Box::new(e) as Box<dyn Encoder>, "software"))
+        }
+        // The auto arm lives in the resolver now (a CUDA frame can only be consumed by
+        // NVENC; else the shared manual-preference/NVIDIA-presence decision).
+        None => anyhow::bail!(
+            "unknown PUNKTFUNK_ENCODER={pref:?} — use auto (default), nvenc, vaapi, vulkan, pyrowave, or software"
+        ),
+    }
+}
+
 /// Open the platform encoder backend. Returns the encoder together with the display label of the
 /// branch that ACTUALLY opened (`nvenc`/`vaapi`/`vulkan`/`amf`/`qsv`/`software`) — the label feeds
 /// the mgmt API's live-session record, and only the open site knows which internal fallback won
@@ -295,209 +563,19 @@ fn open_video_backend(
    };
    #[cfg(target_os = "linux")]
    {
-        // A NEGOTIATED PyroWave session (client advertised + preferred it, plan §3) routes
-        // straight to that backend — the PUNKTFUNK_ENCODER pref below stays a lab override.
-        if codec == Codec::PyroWave {
-            #[cfg(feature = "pyrowave")]
-            {
-                return pyrowave::PyroWaveEncoder::open(width, height, fps, bitrate_bps, chroma)
-                    .map(|e| (Box::new(e) as Box<dyn Encoder>, "pyrowave"));
-            }
-            #[cfg(not(feature = "pyrowave"))]
-            anyhow::bail!(
-                "session negotiated PyroWave but this host was built without --features \
-                 punktfunk-host/pyrowave (the advertisement bit should not have been set)"
-            );
-        }
-        // Pick the GPU encode backend. NVIDIA → NVENC/CUDA (the original path, unchanged);
-        // AMD/Intel → VAAPI (one libavcodec backend for both). Auto-detect by default so a single
-        // Linux binary serves any GPU; `PUNKTFUNK_ENCODER` forces a specific backend (and surfaces
-        // its errors crisply instead of silently trying the other).
-        let pref = pf_host_config::config().encoder_pref.as_str();
-        // AMD/Intel opener. Default = libav VAAPI. With `--features vulkan-encode` +
-        // PUNKTFUNK_VULKAN_ENCODE, an HEVC session instead opens the raw Vulkan Video backend (real
-        // RFI loss recovery the VAAPI path can't express); a failed open falls back to VAAPI so the
-        // stream never dies over the new path. `format`/`bit_depth`/`chroma` only matter to VAAPI —
-        // the Vulkan backend imports the dmabuf and does its own 8-bit 4:2:0 CSC.
-        let open_amd_intel = || -> Result<(Box<dyn Encoder>, &'static str)> {
-            // An HDR session (10-bit + a PQ/BT.2020 capture format) must skip the Vulkan Video
-            // backend — it hardcodes an 8-bit 4:2:0 BT.709 CSC — and take the libav VAAPI path,
-            // which has the P010/Main10/PQ wiring. SDR sessions keep the Vulkan default.
-            #[cfg(feature = "vulkan-encode")]
-            if matches!(codec, Codec::H265 | Codec::Av1)
-                && vulkan_encode_enabled()
-                && !(bit_depth == 10 && format.is_hdr_rgb10())
-            {
-                match vulkan_video::VulkanVideoEncoder::open(
-                    codec,
-                    format,
-                    width,
-                    height,
-                    fps,
-                    bitrate_bps,
-                    cursor_blend,
-                ) {
-                    Ok(e) => {
-                        tracing::info!(
-                            codec = ?codec,
-                            "Linux Vulkan Video encode (real RFI via DPB reference slots) — \
-                             set PUNKTFUNK_VULKAN_ENCODE=0 for libav VAAPI"
-                        );
-                        return Ok((Box::new(e) as Box<dyn Encoder>, "vulkan"));
-                    }
-                    // Native NV12 (PUNKTFUNK_PIPEWIRE_NV12 capture) has no VAAPI fallback:
-                    // libav's dmabuf lane would import the two-plane buffer as packed RGB
-                    // (silent garbage) and its CPU lane bails per frame — die crisply instead.
-                    Err(e) if format == PixelFormat::Nv12 => {
-                        return Err(e.context(
-                            "Vulkan Video open failed on a native-NV12 capture \
-                             — no VAAPI fallback exists; set PUNKTFUNK_PIPEWIRE_NV12=0 to \
-                             restore the packed-RGB negotiation",
-                        ));
-                    }
-                    Err(e) => tracing::warn!(
-                        error = %format!("{e:#}"),
-                        "Vulkan Video encode open failed — falling back to libav VAAPI"
-                    ),
-                }
-            }
-            // Same rule when the Vulkan backend was never eligible (H264 session,
-            // PUNKTFUNK_VULKAN_ENCODE=0, or a build without the feature).
-            if format == PixelFormat::Nv12 {
-                anyhow::bail!(
-                    "native NV12 capture requires the Vulkan Video encoder (HEVC/AV1 \
-                     session, --features vulkan-encode, PUNKTFUNK_VULKAN_ENCODE not 0) — this \
-                     session resolved to libav VAAPI; set PUNKTFUNK_PIPEWIRE_NV12=0 to restore \
-                     the packed-RGB negotiation"
-                );
-            }
-            vaapi::VaapiEncoder::open(
-                codec,
-                format,
-                width,
-                height,
-                fps,
-                bitrate_bps,
-                bit_depth,
-                chroma,
-            )
-            .map(|e| (Box::new(e) as Box<dyn Encoder>, "vaapi"))
-        };
-        let open_nvidia = || -> Result<(Box<dyn Encoder>, &'static str)> {
-            open_nvenc_probed(
-                codec,
-                format,
-                width,
-                height,
-                fps,
-                bitrate_bps,
-                cuda,
-                bit_depth,
-                chroma,
-                cursor_blend,
-            )
-            .map(|e| (e, "nvenc"))
-        };
-        match pref {
-            "nvenc" | "nvidia" | "cuda" => open_nvidia(),
-            "vaapi" | "amd" | "intel" => open_amd_intel(),
-            // Force the raw Vulkan Video HEVC backend (real RFI). Needs `--features vulkan-encode`.
-            "vulkan" | "vulkan-video" => {
-                #[cfg(feature = "vulkan-encode")]
-                {
-                    if !matches!(codec, Codec::H265 | Codec::Av1) {
-                        anyhow::bail!(
-                            "the Vulkan Video encoder supports HEVC + AV1; the session negotiated {codec:?}"
-                        );
-                    }
-                    vulkan_video::VulkanVideoEncoder::open(
-                        codec,
-                        format,
-                        width,
-                        height,
-                        fps,
-                        bitrate_bps,
-                        cursor_blend,
-                    )
-                    .map(|e| (Box::new(e) as Box<dyn Encoder>, "vulkan"))
-                }
-                #[cfg(not(feature = "vulkan-encode"))]
-                {
-                    let _ = (format, bit_depth, chroma);
-                    anyhow::bail!(
-                        "PUNKTFUNK_ENCODER=vulkan requires a build with --features vulkan-encode"
-                    )
-                }
-            }
-            // PyroWave — the opt-in wired-LAN intra-only wavelet codec. Explicit-only, and
-            // EXPERIMENTAL until CODEC_PYROWAVE negotiation lands (plan Phase 2): no shipping
-            // client can decode the stream yet, so this arm exists for host-side bring-up and
-            // latency work only. Vendor-agnostic (any Vulkan 1.3 GPU); ignores the negotiated
-            // codec — every AU is an independently-decodable wavelet frame.
-            "pyrowave" => {
-                #[cfg(feature = "pyrowave")]
-                {
-                    tracing::warn!(
-                        ?codec,
-                        "PUNKTFUNK_ENCODER=pyrowave forces the all-intra wavelet stream \
-                         regardless of the negotiated codec — only a pyrowave-feature client \
-                         that ALSO preferred CODEC_PYROWAVE can display it (lab override; \
-                         normal sessions negotiate it instead)"
-                    );
-                    // The lab override forces the wavelet stream onto a session negotiated for
-                    // another codec — that session's chroma may be HEVC-4:4:4, which the
-                    // pyrowave encoder doesn't do yet, so pin the override to 4:2:0.
-                    pyrowave::PyroWaveEncoder::open(
-                        width,
-                        height,
-                        fps,
-                        bitrate_bps,
-                        ChromaFormat::Yuv420,
-                    )
-                    .map(|e| (Box::new(e) as Box<dyn Encoder>, "pyrowave"))
-                }
-                #[cfg(not(feature = "pyrowave"))]
-                {
-                    anyhow::bail!(
-                        "PUNKTFUNK_ENCODER=pyrowave requires a build with --features punktfunk-host/pyrowave"
-                    )
-                }
-            }
-            // GPU-less software H.264 (openh264) — for a headless / GPU-lost box. Explicit-only:
-            // `auto` never picks it (a box with `/dev/nvidiactl` present but a dead driver would
-            // otherwise wrongly resolve to NVENC). Needs H.264 (openh264 emits only that) and a CPU
-            // RGB frame, which the capturer delivers because the software backend resolves `gpu=false`.
-            "software" | "sw" | "openh264" => {
-                if codec != Codec::H264 {
-                    anyhow::bail!(
-                        "the software encoder emits H.264 only; the session negotiated {codec:?} \
-                         (a client must advertise CODEC_H264 to reach a software host)"
-                    );
-                }
-                let _ = (cuda, bit_depth); // software path is CPU + 8-bit only
-                sw::OpenH264Encoder::open(
-                    format,
-                    width,
-                    height,
-                    fps,
-                    bitrate_bps.min(SW_BITRATE_CEIL),
-                )
-                .map(|e| (Box::new(e) as Box<dyn Encoder>, "software"))
-            }
-            "auto" | "" => {
-                // A CUDA frame can ONLY be consumed by NVENC. Otherwise the shared auto decision
-                // (manual web-console GPU preference, else the NVIDIA-presence probe) picks the
-                // backend — see `linux_auto_is_vaapi`.
-                if cuda || !linux_auto_is_vaapi() {
-                    open_nvidia()
-                } else {
-                    open_amd_intel()
-                }
-            }
-            other => anyhow::bail!(
-                "unknown PUNKTFUNK_ENCODER={other:?} — use auto (default), nvenc, vaapi, vulkan, pyrowave, or software"
-            ),
-        }
+        open_video_backend_linux(
+            pf_host_config::config().encoder_pref.as_str(),
+            codec,
+            format,
+            width,
+            height,
+            fps,
+            bitrate_bps,
+            cuda,
+            bit_depth,
+            chroma,
+            cursor_blend,
+        )
    }
    #[cfg(target_os = "windows")]
    {
@@ -611,8 +689,17 @@ fn open_video_backend(
                // Phase 4 after the field-silence gate.
                #[cfg(feature = "qsv")]
                {
-                    let ffmpeg_forced = std::env::var("PUNKTFUNK_QSV_FFMPEG")
-                        .is_ok_and(|v| v == "1" || v.eq_ignore_ascii_case("true"));
+                    // Trimmed like every sibling knob, so `"1 "` from a shell script or a `.env`
+                    // line takes effect instead of silently doing nothing. Case-insensitivity is
+                    // kept deliberately — this knob already accepted `TRUE` via
+                    // `eq_ignore_ascii_case`, and the bare house `matches!` would have dropped
+                    // that spelling; widening to `yes`/`on` only adds accepted values.
+                    let ffmpeg_forced = std::env::var("PUNKTFUNK_QSV_FFMPEG").is_ok_and(|v| {
+                        matches!(
+                            v.trim().to_ascii_lowercase().as_str(),
+                            "1" | "true" | "yes" | "on"
+                        )
+                    });
                    if !ffmpeg_forced {
                        match qsv::QsvEncoder::open(
                            codec,
@@ -853,18 +940,31 @@ fn vulkan_encode_enabled() -> bool {
 /// negotiation (`OutputFormat::nv12_native` → `ZeroCopyPolicy::native_nv12_session`), which
 /// then PREFERS gamescope's producer-side NV12 pod (default-on; `PUNKTFUNK_PIPEWIRE_NV12=0`
 /// escapes at the capture gate).
+///
+/// This verdict is load-bearing in a way the other capability helpers are not: once the producer
+/// has been asked for two-plane NV12 there is **no fallback**. [`open_video`] deliberately makes a
+/// failed Vulkan open FATAL for an NV12 capture rather than degrading to libav VAAPI, because VAAPI
+/// would import that buffer as packed RGB and stream silent garbage. So a wrong `true` here does
+/// not cost quality — it kills the session at its first frame. Hence the real device probe below,
+/// and hence the conjuncts are ordered cheapest-first so it only runs when everything else already
+/// said yes.
 #[cfg(target_os = "linux")]
 pub fn linux_native_nv12_ok(codec: Codec) -> bool {
    #[cfg(feature = "vulkan-encode")]
    {
        matches!(codec, Codec::H265 | Codec::Av1)
            && vulkan_encode_enabled()
-            // NVENC/PyroWave prefs never open the Vulkan Video backend; every other pref
-            // (auto/vaapi/amd/intel/vulkan) tries it first on AMD/Intel — see [`open_video`].
-            && !matches!(
-                pf_host_config::config().encoder_pref.as_str(),
-                "nvenc" | "nvidia" | "cuda" | "pyrowave"
-            )
+            // Which backend this host actually resolves to. This used to be a denylist of the
+            // EXPLICIT prefs that skip Vulkan Video ("nvenc"|"nvidia"|"cuda"|"pyrowave"), which
+            // silently missed the one that matters: the DEFAULT `encoder_pref` is `""`, and `""`
+            // resolves to `auto`, which on an NVIDIA box opens NVENC. So a stock NVIDIA host passed
+            // this gate. `linux_zero_copy_is_vaapi` layers the pref on top of the same auto decision
+            // `open_video` makes, which is exactly what the note on `linux_auto_is_vaapi` says a
+            // capability probe must consult — and it is what the downstream consumer of this very
+            // verdict already uses to pick its zero-copy path.
+            && linux_zero_copy_is_vaapi()
+            // …and only then ask the GPU. Ordered last on purpose: it opens a Vulkan instance.
+            && vulkan_encode_available(codec)
    }
    #[cfg(not(feature = "vulkan-encode"))]
    {
@@ -873,6 +973,42 @@ pub fn linux_native_nv12_ok(codec: Codec) -> bool {
    }
 }

+/// Can this GPU + driver actually open a Vulkan Video **encode** session for `codec`? Cached per
+/// (selected GPU, codec) — the [`can_encode_10bit`] idiom, with the probe run outside the lock.
+///
+/// Only [`linux_native_nv12_ok`] consults this, and only for the no-fallback decision described
+/// there. It is deliberately NOT wired into [`open_video`]'s dispatch: that path already degrades
+/// to VAAPI on a failed open for every non-NV12 capture, so making it pay for a probe would buy
+/// nothing. Prediction and truth stay separate — the probe answers "would it open", the open itself
+/// reports what actually did.
+#[cfg(all(target_os = "linux", feature = "vulkan-encode"))]
+fn vulkan_encode_available(codec: Codec) -> bool {
+    use std::collections::HashMap;
+    use std::sync::{Mutex, OnceLock};
+    static CACHE: OnceLock<Mutex<HashMap<(String, &'static str), bool>>> = OnceLock::new();
+    let key = (pf_gpu::selection_key(), codec.label());
+    let cache = CACHE.get_or_init(|| Mutex::new(HashMap::new()));
+    if let Some(v) = cache.lock().unwrap().get(&key) {
+        return *v;
+    }
+    let verdict = vulkan_video::probe_encode_support(codec);
+    let ok = verdict.is_ok();
+    match &verdict {
+        Ok(()) => tracing::info!(
+            ?codec,
+            "Vulkan Video encode probed OK — producer-native NV12 capture is eligible"
+        ),
+        Err(why) => tracing::info!(
+            ?codec,
+            why = *why,
+            "Vulkan Video encode unavailable — keeping the packed-RGB capture negotiation \
+             (the native-NV12 path has no VAAPI fallback)"
+        ),
+    }
+    cache.lock().unwrap().insert(key, ok);
+    ok
+}
+
 /// Cheap, side-effect-free NVIDIA-presence probe for the `auto` backend selector: the NVIDIA
 /// kernel driver exposes these device nodes, AMD/Intel boxes have neither. Deliberately does NOT
 /// create a CUDA context (that would allocate GPU state on every host that merely *might* be
@@ -904,6 +1040,67 @@ fn linux_auto_is_vaapi() -> bool {
    !nvidia_present()
 }

+/// The resolved Linux encode backend — the Linux twin of [`WindowsBackend`] (WP7.6). ONE alias
+/// table, consumed by BOTH [`open_video_backend`]'s dispatch and every capability/advertisement
+/// mirror below, so the two can never drift again (the audit counted five partial hand-copies of
+/// this decision). Labels still come from the open sites (the mgmt-record convention) — this
+/// resolves which ARM runs, never what actually opened.
+#[cfg(target_os = "linux")]
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+enum LinuxBackend {
+    Nvenc,
+    AmdIntel,
+    Vulkan,
+    Pyrowave,
+    Software,
+}
+
+/// The pure core. `None` = the pref names no backend — [`open_video_backend`] BAILS on that,
+/// while the capability mirrors historically resolved it as auto (via
+/// [`linux_resolved_backend`]); that divergence is recorded, not accidental. `auto_is_vaapi` is
+/// deliberately LAZY: explicit prefs must stay zero-probe — `/serverinfo` polls through the
+/// mirrors, and `gamestream/serverinfo.rs` already litigated exactly that probe cost.
+#[cfg(target_os = "linux")]
+fn resolve_linux_backend(
+    pref: &str,
+    auto_is_vaapi: impl FnOnce() -> bool,
+    cuda: bool,
+) -> Option<LinuxBackend> {
+    Some(match pref {
+        "nvenc" | "nvidia" | "cuda" => LinuxBackend::Nvenc,
+        "vaapi" | "amd" | "intel" => LinuxBackend::AmdIntel,
+        "vulkan" | "vulkan-video" => LinuxBackend::Vulkan,
+        "pyrowave" => LinuxBackend::Pyrowave,
+        "software" | "sw" | "openh264" => LinuxBackend::Software,
+        // A CUDA frame can ONLY be consumed by NVENC. Otherwise the shared auto decision
+        // (manual web-console GPU preference, else the NVIDIA-presence probe) picks the backend —
+        // see `linux_auto_is_vaapi`.
+        "auto" | "" => {
+            if cuda || !auto_is_vaapi() {
+                LinuxBackend::Nvenc
+            } else {
+                LinuxBackend::AmdIntel
+            }
+        }
+        _ => return None,
+    })
+}
+
+/// Config-reading wrapper for the capability/advertisement mirrors: an unknown pref resolves as
+/// auto — exactly what every mirror's old `_` arm did. `cuda = false`: the mirrors answer
+/// pre-session questions (a CUDA payload exists only inside a live NVIDIA session).
+#[cfg(target_os = "linux")]
+fn linux_resolved_backend() -> LinuxBackend {
+    let pref = pf_host_config::config().encoder_pref.as_str();
+    resolve_linux_backend(pref, linux_auto_is_vaapi, false).unwrap_or_else(|| {
+        if linux_auto_is_vaapi() {
+            LinuxBackend::AmdIntel
+        } else {
+            LinuxBackend::Nvenc
+        }
+    })
+}
+
 /// The dmabuf modifiers the PyroWave encoder's Vulkan device imports for the capture's
 /// packed-RGB fourcc — advertised by the capture when the pyrowave passthrough is active
 /// (the VAAPI LINEAR-only policy starves it on Mutter+NVIDIA, which allocates tiled only).
@@ -917,13 +1114,25 @@ pub fn pyrowave_capture_modifiers(fourcc: u32) -> Vec<u64> {
 /// passthrough for VAAPI vs the EGL→CUDA import for NVENC).
 #[cfg(target_os = "linux")]
 pub fn linux_zero_copy_is_vaapi() -> bool {
-    match pf_host_config::config().encoder_pref.as_str() {
-        "nvenc" | "nvidia" | "cuda" => false,
-        "vaapi" | "amd" | "intel" => true,
+    linux_zero_copy_is_vaapi_for(linux_resolved_backend())
+}
+/// The zero-copy-plane decision for an ALREADY-resolved backend — so a caller that has just
+/// resolved (e.g. `host_wire_caps`, which consults several gates per call on a POLLED endpoint)
+/// pays for the resolution once instead of once per gate (the `serverinfo.rs` probe-cost class,
+/// caught by the WP7.6 diff critic).
+#[cfg(target_os = "linux")]
+fn linux_zero_copy_is_vaapi_for(backend: LinuxBackend) -> bool {
+    match backend {
+        LinuxBackend::Nvenc => false,
+        LinuxBackend::AmdIntel => true,
        // PyroWave ingests the raw capture dmabuf itself (Vulkan import + compute CSC) on ANY
        // vendor — it must get the passthrough payload, never the EGL→CUDA import.
-        "pyrowave" => true,
-        _ => linux_auto_is_vaapi(),
+        LinuxBackend::Pyrowave => true,
+        // EXACTLY the old `_` fallthrough for these two prefs — preserved, not endorsed. The
+        // vulkan-on-NVIDIA half is FILED (EGL→CUDA capture negotiated for a dmabuf-importing
+        // backend; reachable only via the explicit lab pref); the software half is latent (a
+        // software host pins H.264, so the GPU-backend probes this steers never negotiate).
+        LinuxBackend::Vulkan | LinuxBackend::Software => linux_auto_is_vaapi(),
    }
 }

@@ -1216,9 +1425,17 @@ pub fn windows_resolved_backend() -> WindowsBackend {
 pub fn resolved_backend_is_gpu() -> bool {
    !matches!(windows_resolved_backend(), WindowsBackend::Software)
 }
-/// Linux/other: every backend but the GPU-less software encoder (openh264) is GPU-resident. Config-backed
-/// (mirrors `session_plan::resolve_encoder`; the NVENC vs VAAPI split is auto-detected in [`open_video`]).
-#[cfg(not(target_os = "windows"))]
+/// Linux: every backend but the GPU-less software encoder (openh264) is GPU-resident. Resolver-
+/// backed (WP7.6); cross-crate mirrors still exist in `session_plan::resolve_encoder` and
+/// `gamestream/serverinfo.rs` — out of this crate's reach, listed in the WP7.6 design doc.
+#[cfg(target_os = "linux")]
+pub fn resolved_backend_is_gpu() -> bool {
+    linux_resolved_backend() != LinuxBackend::Software
+}
+/// Other non-Windows targets (the macOS dev host): keep the string gate — no resolver exists
+/// there. `not(any(...))`, NEVER a target list, so no exotic target loses the fn (the
+/// `can_encode_444`/`can_encode_10bit` shape).
+#[cfg(not(any(target_os = "windows", target_os = "linux")))]
 pub fn resolved_backend_is_gpu() -> bool {
    !matches!(
        pf_host_config::config().encoder_pref.as_str(),
@@ -1393,7 +1610,9 @@ mod nvenc;
 #[path = "enc/linux/nvenc_cuda.rs"]
 mod nvenc_cuda;
 // Actionable `NVENCSTATUS` → cause mapping shared by both direct-NVENC backends, so a failed
-// session open logs "update/reboot the driver" instead of the old misleading "(no NVIDIA GPU?)".
+// session open names its real cause instead of the old misleading "(no NVIDIA GPU?)" — including
+// splitting the two opposite failures the driver reports as the SAME `INVALID_VERSION` (a genuine
+// driver skew vs. this process's driver state going bad after a session already opened).
 #[cfg(all(any(target_os = "linux", target_os = "windows"), feature = "nvenc"))]
 #[path = "enc/nvenc_status.rs"]
 mod nvenc_status;
@@ -1402,6 +1621,20 @@ mod nvenc_status;
 #[cfg(all(any(target_os = "linux", target_os = "windows"), feature = "nvenc"))]
 #[path = "enc/nvenc_core.rs"]
 mod nvenc_core;
+// The slot-family RFI recovery policy (WP7.2) shared by native AMF, native QSV and Vulkan Video —
+// the taint sweep + pre-loss anchor pick extracted from three hand-copies that had already
+// diverged once (the fecbec2d sweep predates the Vulkan backend's carve-out and was never ported
+// to it until a later fix). Policy only: every mechanism (LTR bitfield, RefListCtrl, DPB slot
+// table) stays in its backend. cfg = the union of the callers': amf.rs is featureless on Windows,
+// vulkan_video needs `vulkan-encode` on Linux — and each ITEM inside is live under the module's
+// whole cfg because `plan_slot_recovery` delegates to `pick_anchor` (dead_code is an item lint;
+// module-granular reasoning is how a Windows-dead helper reaches main).
+#[cfg(any(
+    target_os = "windows",
+    all(target_os = "linux", feature = "vulkan-encode")
+))]
+#[path = "enc/rfi.rs"]
+mod rfi;
 // Shared libavcodec glue (`pixel_to_av`, swscale consts) for the three libav backends — Linux
 // NVENC + VAAPI and Windows AMF/QSV — so the byte-identical pieces live once (plan §2.2, Tier 2).
 #[cfg(any(target_os = "linux", all(target_os = "windows", feature = "amf-qsv")))]
@@ -1504,4 +1737,153 @@ mod tests {
        };
        assert_eq!(none.wire_mask(), None);
    }
+
+    /// WP7.7 guard (the cheap half): every `Encoder` trait method must be explicitly forwarded by
+    /// `TrackedEncoder`. A defaulted trait method that isn't forwarded silently no-ops through the
+    /// wrapper — the trap has bitten three times (`set_wire_chunking`'s §4.4 chunking probe,
+    /// `set_pipelined`'s LN3 escalation, `applied_bitrate_bps`'s ABR truth), and every Phase 7
+    /// consolidation that adds a trait method re-arms it. Source-text parse, not reflection: both
+    /// blocks are top-level rustfmt items, so each ends at the first column-0 `}` and every method
+    /// name sits on a line starting `fn ` (a wrapped signature still carries the name there).
+    #[test]
+    fn tracked_encoder_forwards_every_trait_method() {
+        fn item_block<'a>(src: &'a str, marker: &str) -> &'a str {
+            let start = src
+                .find(marker)
+                .unwrap_or_else(|| panic!("marker {marker:?} not found — update this guard"));
+            let body = &src[start..];
+            let end = body
+                .find("\n}")
+                .unwrap_or_else(|| panic!("no column-0 close brace after {marker:?}"));
+            &body[..end]
+        }
+        fn fn_names(block: &str) -> std::collections::BTreeSet<&str> {
+            block
+                .lines()
+                .map(str::trim_start)
+                .filter(|l| !l.starts_with("//"))
+                .filter_map(|l| l.strip_prefix("fn "))
+                .map(|rest| {
+                    rest.split(|c: char| !(c.is_ascii_alphanumeric() || c == '_'))
+                        .next()
+                        .expect("split yields at least one item")
+                })
+                .collect()
+        }
+        // `find` takes the FIRST occurrence: the real impl precedes this test's own copy of the
+        // marker string in the included source.
+        let trait_fns = fn_names(item_block(
+            include_str!("enc/codec.rs"),
+            "pub trait Encoder: Send {",
+        ));
+        let impl_fns = fn_names(item_block(
+            include_str!("lib.rs"),
+            "impl Encoder for TrackedEncoder {",
+        ));
+        assert!(
+            trait_fns.len() >= 12,
+            "only {} trait methods parsed — the extraction markers have rotted, fix the parse \
+             before trusting this guard",
+            trait_fns.len()
+        );
+        let missing: Vec<_> = trait_fns.difference(&impl_fns).collect();
+        assert!(
+            missing.is_empty(),
+            "Encoder methods NOT forwarded by TrackedEncoder: {missing:?} — the host loop only \
+             ever holds the wrapped box, so an unforwarded default silently disables the feature \
+             for every session. Forward each one in `impl Encoder for TrackedEncoder`."
+        );
+        // The reverse direction (impl fn absent from the trait) is a compile error, so equality
+        // here is pure belt-and-braces against a parse regression.
+        assert_eq!(trait_fns, impl_fns);
+    }
+
+    /// WP7.6: the resolver's full alias table, pinned. The panicking closure is the laziness
+    /// contract — an explicit pref must resolve WITHOUT running the auto probe (`/serverinfo`
+    /// polls through the mirrors; `gamestream/serverinfo.rs` litigated exactly that cost).
+    #[cfg(target_os = "linux")]
+    #[test]
+    fn linux_backend_resolver_table() {
+        use LinuxBackend::*;
+        let no_probe = || -> bool { panic!("explicit prefs must not run the auto probe") };
+        for pref in ["nvenc", "nvidia", "cuda"] {
+            assert_eq!(resolve_linux_backend(pref, no_probe, false), Some(Nvenc));
+        }
+        for pref in ["vaapi", "amd", "intel"] {
+            assert_eq!(resolve_linux_backend(pref, no_probe, false), Some(AmdIntel));
+        }
+        for pref in ["vulkan", "vulkan-video"] {
+            assert_eq!(resolve_linux_backend(pref, no_probe, false), Some(Vulkan));
+        }
+        assert_eq!(
+            resolve_linux_backend("pyrowave", no_probe, false),
+            Some(Pyrowave)
+        );
+        for pref in ["software", "sw", "openh264"] {
+            assert_eq!(resolve_linux_backend(pref, no_probe, false), Some(Software));
+        }
+        // auto (and the default empty pref): the probe decides…
+        assert_eq!(resolve_linux_backend("", || true, false), Some(AmdIntel));
+        assert_eq!(resolve_linux_backend("auto", || false, false), Some(Nvenc));
+        // …except a CUDA frame, which only NVENC can consume — and that short-circuits the
+        // probe too (`||` order).
+        assert_eq!(resolve_linux_backend("auto", no_probe, true), Some(Nvenc));
+        // Unknown prefs name no backend: the dispatch bails; the mirrors' wrapper maps this to
+        // auto (the historical `_`-arm behavior, recorded in the design doc).
+        assert_eq!(resolve_linux_backend("banana", no_probe, false), None);
+        // An explicit pref is never overridden by `cuda` (a CUDA payload cannot appear in a
+        // session whose pref forced a non-NVENC backend; the table must not mask such a bug).
+        assert_eq!(
+            resolve_linux_backend("vaapi", no_probe, true),
+            Some(AmdIntel)
+        );
+    }
+
+    /// WP7.6: the REAL dispatch through the resolver, GPU-free via the software arm (openh264 is
+    /// vendored + CPU-only; its own tests already run on every leg). This is the only in-CI
+    /// execution of the Linux dispatch — before this test it had zero test call sites. The pref
+    /// is INJECTED through `open_video_backend_linux` — the first draft mutated `PUNKTFUNK_ENCODER`
+    /// via `set_var`, the binary's first non-ignored env mutation, racing `getenv` in the sw/codec/
+    /// nvenc_core tests (diff-critic catch); the seam removes the race and the config-latch
+    /// coupling outright.
+    #[cfg(target_os = "linux")]
+    #[test]
+    fn open_video_backend_dispatches_software() {
+        let (enc, label) = open_video_backend_linux(
+            "software",
+            Codec::H264,
+            PixelFormat::Bgrx,
+            64,
+            64,
+            30,
+            1_000_000,
+            false,
+            8,
+            ChromaFormat::Yuv420,
+            false,
+        )
+        .expect("software arm must open GPU-free");
+        assert_eq!(label, "software");
+        drop(enc);
+        // The software arm is codec-gated — a non-H.264 session must take the bail, not fall
+        // through to another backend.
+        let err = match open_video_backend_linux(
+            "software",
+            Codec::H265,
+            PixelFormat::Bgrx,
+            64,
+            64,
+            30,
+            1_000_000,
+            false,
+            8,
+            ChromaFormat::Yuv420,
+            false,
+        ) {
+            // `expect_err` needs `Ok: Debug` and `Box<dyn Encoder>` isn't — match instead.
+            Ok(_) => panic!("software emits H.264 only; an H.265 session must be refused"),
+            Err(e) => e,
+        };
+        assert!(err.to_string().contains("H.264"), "{err:#}");
+    }
 }
@@ -360,6 +360,28 @@ fn run_inner(mut opts: SessionOpts, mut mode: ModeCtl) -> Result<Option<Outcome>
    // identity and the session window gets the default-Wayland icon (the Linux analog of
    // the AppUserModelID adoption above).
    sdl3::hint::set("SDL_APP_ID", "io.unom.Punktfunk");
+    // `PUNKTFUNK_DRM_CARD=<n>` → SDL's KMSDRM device index, for a compositor-less (kiosk/embedded)
+    // run. SDL enumerates /dev/dri/card* and takes the first one it can open, which on a
+    // multi-GPU box is regularly the WRONG one: measured on a two-card machine it chose the card
+    // a live compositor already held DRM master on and failed at swapchain creation, while the
+    // idle card with the connected display sat unused. There is no reliable way to pick from
+    // inside the process (detecting "already mastered" needs the ioctl that taking master IS), so
+    // this stays an explicit operator choice rather than fragile auto-detection.
+    // Ignored unless the kmsdrm backend is actually in use.
+    if let Ok(card) = std::env::var("PUNKTFUNK_DRM_CARD") {
+        if card.chars().all(|c| c.is_ascii_digit()) && !card.is_empty() {
+            tracing::info!(
+                card,
+                "PUNKTFUNK_DRM_CARD: pinning SDL's KMSDRM device index"
+            );
+            sdl3::hint::set("SDL_KMSDRM_DEVICE_INDEX", &card);
+        } else {
+            tracing::warn!(
+                card,
+                "PUNKTFUNK_DRM_CARD must be a card NUMBER (e.g. 0) — ignoring"
+            );
+        }
+    }
    let sdl = sdl3::init().context("SDL init")?;
    let video = sdl.video().context("SDL video")?;
    let events = sdl.event().context("SDL events")?;
@@ -3,9 +3,34 @@
 use super::setup::pick_formats;
 use super::{OverlayPipe, Presenter};
 use crate::csc::CscPass;
-use anyhow::{Context as _, Result};
+use anyhow::{anyhow, Context as _, Result};
 use ash::vk;

+/// Extra guidance appended to a swapchain-creation failure when SDL is on the KMSDRM backend —
+/// i.e. a compositor-less kiosk/embedded run, where the bare Vulkan error is close to useless.
+///
+/// Measured on an NVIDIA proprietary + KMSDRM box: SDL opens the card, Vulkan enumerates the GPU
+/// *and* the display (`VK_KHR_display` reports the connected HDMI connector), then
+/// `vkCreateSwapchainKHR` returns ERROR_INITIALIZATION_FAILED — as root, with `nvidia_drm.modeset`
+/// on, on a card no compositor was using. So it is neither permissions nor DRM master: NVIDIA's
+/// direct-to-display path wants the display leased (`vkAcquireDrmDisplayEXT`) and SDL's KMSDRM
+/// surface path does not do that for it. Empty on every other backend, where the message would be
+/// noise.
+fn kmsdrm_swapchain_hint() -> String {
+    let kmsdrm = std::env::var("SDL_VIDEODRIVER").is_ok_and(|v| v.eq_ignore_ascii_case("kmsdrm"));
+    if !kmsdrm {
+        return String::new();
+    }
+    let card = std::env::var("PUNKTFUNK_DRM_CARD").unwrap_or_else(|_| "unset".into());
+    format!(
+        " — under SDL_VIDEODRIVER=kmsdrm (PUNKTFUNK_DRM_CARD={card}). Check, in order: the card \
+         has a CONNECTED connector (`cat /sys/class/drm/card*-*/status`); nothing else holds DRM \
+         master on it (a running compositor does — pin another card with PUNKTFUNK_DRM_CARD=<n>); \
+         and the driver is Mesa. NVIDIA's proprietary direct-display path is known to fail here \
+         even as root with a display Vulkan can enumerate."
+    )
+}
+
 impl Presenter {
    /// (Re)build the swapchain for the window's current pixel size. Also the resize path.
    pub fn recreate_swapchain(&mut self, window: &sdl3::video::Window) -> Result<()> {
@@ -66,8 +91,8 @@ impl Presenter {
            .present_mode(self.present_mode)
            .clipped(true)
            .old_swapchain(old);
-        let swapchain =
-            unsafe { self.swap_d.create_swapchain(&info, None) }.context("vkCreateSwapchainKHR")?;
+        let swapchain = unsafe { self.swap_d.create_swapchain(&info, None) }
+            .map_err(|e| anyhow!("vkCreateSwapchainKHR: {e}{}", kmsdrm_swapchain_hint()))?;
        // The old swapchain and everything tied to its images dies NOW: the fence
        // quiesce covered our own command buffers, the queue drain above covered the
        // presentation engine's semaphore waits — nothing can still reference them.
@@ -11,7 +11,7 @@ use crate::dmabuf;
 use anyhow::{anyhow, bail, Context as _, Result};
 use ash::vk;
 use ash::vk::Handle as _;
-use std::ffi::CString;
+use std::ffi::{c_char, CString};

 impl Presenter {
    /// Bring up instance → surface → device → swapchain over an SDL window.
@@ -40,7 +40,10 @@ impl Presenter {
            .iter()
            .map(|e| CString::new(e.as_str()).unwrap())
            .collect();
-        let ext_ptrs: Vec<*const i8> = ext_cstrings.iter().map(|e| e.as_ptr()).collect();
+        // `c_char`, not `i8`: plain `char` is SIGNED on x86_64 but UNSIGNED on aarch64, so a
+        // hardcoded `*const i8` compiles on the desktop targets and fails to match ash's
+        // `&[*const c_char]` on ARM.
+        let ext_ptrs: Vec<*const c_char> = ext_cstrings.iter().map(|e| e.as_ptr()).collect();
        let instance = unsafe {
            entry.create_instance(
                &vk::InstanceCreateInfo::default()
@@ -1139,18 +1139,67 @@ fn dm_survives_masked_unit(dm: &str) -> bool {
    dm == "sddm.service"
 }

-/// Stop the display manager for a takeover on a mask-fragile DM flavor. Plain `systemctl stop` on
-/// the SYSTEM bus — succeeds as root or under an operator polkit rule scoped to the DM unit (see
-/// docs); fails cleanly otherwise ("interactive authentication required"), in which case the
-/// caller degrades to attach.
-fn try_stop_display_manager(dm: &str) -> bool {
-    Command::new("systemctl")
-        .args(["stop", dm])
+/// The packaged privileged fallback for the display-manager takeover verbs: a root helper behind
+/// its own polkit action (`io.unom.punktfunk.dm-helper`, `allow_any` — the mechanism these
+/// distros use for their own session switcher, e.g. Nobara's `os-session-select`), so the managed
+/// takeover works out of the box on mask-fragile DM flavors with no hand-installed polkit rule.
+/// The helper derives the DM unit from the `display-manager.service` symlink itself, so this
+/// process never gets to name an arbitrary unit across the privilege boundary. Two layouts: the
+/// rpm/deb `libexec` path (what the shipped policy annotates) and Arch's `/usr/lib/<pkg>` (its
+/// PKGBUILD rewrites the annotation to match).
+const DM_HELPER_PATHS: &[&str] = &[
+    "/usr/libexec/punktfunk/pf-dm-helper",
+    "/usr/lib/punktfunk/pf-dm-helper",
+];
+
+/// Run the packaged DM helper (`stop` | `restore`) via pkexec. `false` when the helper isn't
+/// installed (tarball/old package), pkexec is missing, or polkit denies the action.
+fn dm_helper(verb: &str) -> bool {
+    let Some(helper) = DM_HELPER_PATHS
+        .iter()
+        .find(|p| std::path::Path::new(p).exists())
+    else {
+        return false;
+    };
+    Command::new("pkexec")
+        .arg(helper)
+        .arg(verb)
        .status()
        .map(|s| s.success())
        .unwrap_or(false)
 }

+/// Stop the display manager for a takeover on a mask-fragile DM flavor. Plain `systemctl stop` on
+/// the SYSTEM bus first — succeeds as root or under an operator polkit rule scoped to the DM unit
+/// (see docs); fails cleanly otherwise ("interactive authentication required") — then the
+/// packaged pkexec helper. `false` means no privilege path exists and the caller degrades to
+/// attach.
+fn try_stop_display_manager(dm: &str) -> bool {
+    let direct = Command::new("systemctl")
+        .args(["stop", dm])
+        .status()
+        .map(|s| s.success())
+        .unwrap_or(false);
+    direct || dm_helper("stop")
+}
+
+/// Restore the display manager: `reset-failed` (a relogin loop may have tripped the unit's start
+/// limit, and a plain restart is refused until the accounting clears) + `restart` — its autologin
+/// session Exec brings the box's own session back up. Plain system-bus verbs first (root / an
+/// operator polkit rule), then the packaged pkexec helper, whose `restore` verb performs the same
+/// two steps as root.
+fn restore_display_manager(dm: &str) -> bool {
+    let _ = Command::new("systemctl")
+        .args(["reset-failed", dm])
+        .status();
+    let direct = Command::new("systemctl")
+        .args(["restart", dm])
+        .status()
+        .map(|s| s.success())
+        .unwrap_or(false);
+    direct || dm_helper("restore")
+}
+
 /// The distro's session-switch helper (ChimeraOS/Nobara layout). Its USER pass records the
 /// sentinel + self-pkexecs (authorized `allow_any` by the distro's own polkit action policy, so
 /// it works from our sessionless context); its ROOT pass rewrites the DM autologin config — but
@@ -1219,10 +1268,13 @@ fn honor_session_select_switch(dm: String) {
    clear_takeover();
    *MANAGED_SESSION.lock().unwrap_or_else(|e| e.into_inner()) = None;
    stop_session(SESSION_UNIT); // dead already (the switch shut its Steam down) — clear the unit
-    let _ = Command::new("systemctl")
-        .args(["reset-failed", &dm])
-        .status();
-    let _ = Command::new("systemctl").args(["start", &dm]).status();
+    if !restore_display_manager(&dm) {
+        tracing::warn!(
+            %dm,
+            "gamescope: display-manager start was denied — the desktop switch may need a manual \
+             `systemctl restart` of the DM"
+        );
+    }
    let deadline = Instant::now() + Duration::from_secs(10);
    while Instant::now() < deadline {
        let active = Command::new("systemctl")
@@ -1343,8 +1395,10 @@ fn stop_autologin_sessions() -> Result<()> {
        if !try_stop_display_manager(&dm) {
            bail!(
                "the box's gaming session is driven by {dm}, which does not survive a masked \
-                 session unit, and stopping it needs privilege — install the punktfunk \
-                 display-manager polkit rule (see docs) to enable the managed takeover"
+                 session unit, and stopping it needs privilege — the packaged pf-dm-helper \
+                 polkit action is missing or was denied (reinstall the punktfunk package, or \
+                 install the display-manager polkit rule from the docs) so the managed takeover \
+                 is unavailable"
            );
        }
        tracing::info!(
@@ -1668,14 +1722,7 @@ fn do_restore_tv_session() {
    // (`reset-failed` clears the relogin start-limit accounting, then `restart`); its autologin
    // session Exec starts the gamescope unit itself.
    if let Some(dm) = dm {
-        let _ = Command::new("systemctl")
-            .args(["reset-failed", &dm])
-            .status();
-        let restart = Command::new("systemctl")
-            .args(["restart", &dm])
-            .status()
-            .map(|s| s.success())
-            .unwrap_or(false);
+        let restart = restore_display_manager(&dm);
        if restart {
            tracing::info!(%dm, "restored the display manager (its autologin brings gaming mode back)");
        } else if crate::try_recover_session() {
@@ -340,9 +340,19 @@ pub(crate) fn is_available() -> bool {
 /// against PipeWire ≥ 1.6 (a loop-lock bug) and a stuck link head-blocks the whole daemon.
 const MIN_GAMESCOPE: (u32, u32, u32) = (3, 16, 22);

-/// Best-effort: warn loudly if the installed gamescope is older than [`MIN_GAMESCOPE`]. Parsing
-/// failures are silent (don't block a possibly-fine custom build) — this is a diagnostic, not a
-/// gate. Returns the parsed version when it could read one.
+/// First gamescope that paints the Steam overlay (Shift+Tab / Quick Access Menu) into its built-in
+/// PipeWire node. `paint_pipewire()` is a *separate, reduced* composite from the display scanout;
+/// the overlay-window paint (gated on the consumer negotiating `gamescope_focus_appid == 0`, which
+/// we do by never advertising that property — see the capturer's EnumFormat builders) first ships
+/// in 3.16.23 (gamescope commits `ccd62074` + `f8b33d38`). Below this the overlay is *never* in the
+/// node, so it cannot appear in the stream no matter what the host does. The cursor and
+/// external-overlay / notification layers are excluded on *every* version (handled host-side).
+const MIN_GAMESCOPE_OVERLAY: (u32, u32, u32) = (3, 16, 23);
+
+/// Best-effort: warn if the installed gamescope is older than [`MIN_GAMESCOPE`] (capture is
+/// unreliable) or than [`MIN_GAMESCOPE_OVERLAY`] (capture works but the Steam overlay can't reach
+/// the stream). Parsing failures are silent (don't block a possibly-fine custom build) — this is a
+/// diagnostic, not a gate. Returns the parsed version when it could read one.
 pub(super) fn check_gamescope_version() -> Option<(u32, u32, u32)> {
    let out = Command::new("gamescope").arg("--version").output().ok()?;
    // gamescope prints the version banner to stderr on some builds, stdout on others.
@@ -360,6 +370,18 @@ pub(super) fn check_gamescope_version() -> Option<(u32, u32, u32)> {
             capture deadlock against PipeWire ≥ 1.6 (a wedged link head-blocks the daemon); \
             upgrade gamescope or use PUNKTFUNK_COMPOSITOR=kwin|mutter"
        );
+    } else if ver < MIN_GAMESCOPE_OVERLAY {
+        // Capture is fine; the Steam overlay just won't be in the frame gamescope hands us.
+        tracing::warn!(
+            found = %format!("{}.{}.{}", ver.0, ver.1, ver.2),
+            min = %format!(
+                "{}.{}.{}",
+                MIN_GAMESCOPE_OVERLAY.0, MIN_GAMESCOPE_OVERLAY.1, MIN_GAMESCOPE_OVERLAY.2
+            ),
+            "gamescope is older than the first version that paints the Steam overlay (Shift+Tab / \
+             Quick Access Menu) into its PipeWire node — the overlay will be absent from the \
+             stream until you upgrade gamescope (the cursor is composited host-side regardless)"
+        );
    }
    Some(ver)
 }
@@ -379,7 +401,7 @@ fn parse_version(text: &str) -> Option<(u32, u32, u32)> {

 #[cfg(test)]
 mod tests {
-    use super::{parse_version, steam_appid_from_launch, MIN_GAMESCOPE};
+    use super::{parse_version, steam_appid_from_launch, MIN_GAMESCOPE, MIN_GAMESCOPE_OVERLAY};

    #[test]
    fn parses_steam_appid_from_launch() {
@@ -425,4 +447,15 @@ mod tests {
        assert!(parse_version("gamescope version 3.16.22").unwrap() >= MIN_GAMESCOPE);
        assert!(parse_version("gamescope version 3.17.0").unwrap() >= MIN_GAMESCOPE);
    }
+
+    #[test]
+    fn overlay_threshold_brackets_the_fix() {
+        // 3.16.22 captures fine but predates the overlay-in-pipewire paint — it sits in the
+        // "capture works, overlay absent" window `[MIN_GAMESCOPE, MIN_GAMESCOPE_OVERLAY)`, which is
+        // exactly the `else if` warn arm; 3.16.23 is the first to include the overlay.
+        assert!(parse_version("gamescope version 3.16.22").unwrap() >= MIN_GAMESCOPE);
+        assert!(parse_version("gamescope version 3.16.22").unwrap() < MIN_GAMESCOPE_OVERLAY);
+        assert!(parse_version("gamescope version 3.16.23").unwrap() >= MIN_GAMESCOPE_OVERLAY);
+        assert!(parse_version("gamescope version 3.16.25").unwrap() >= MIN_GAMESCOPE_OVERLAY);
+    }
 }
@@ -30,3 +30,7 @@ ash = "0.38"
 # (SCM_RIGHTS carries the fds; pixels never cross the socket).
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
+
+# Linux-only like the code under test: the render-node scan is exercised against a fixture tree.
+[target.'cfg(target_os = "linux")'.dev-dependencies]
+tempfile = "3"
@@ -34,6 +34,62 @@ const EGL_DMA_BUF_PLANE0_MODIFIER_HI_EXT: egl::Attrib = 0x3444;
 mod gl;
 use gl::*;

+/// PCI vendor id of the only GPU this importer can drive — everything below it ends in a CUDA
+/// context. Same constant the Vulkan bridge selects its physical device by.
+const PCI_VENDOR_NVIDIA: u32 = 0x10de;
+
+/// The NVIDIA DRM render node: `PUNKTFUNK_ZEROCOPY_RENDER_NODE` > the first `/dev/dri/renderD*`
+/// whose sysfs PCI vendor is NVIDIA > `/dev/dri/renderD128`.
+///
+/// The fallback used to be the *whole* selection, which quietly assumed NVIDIA owns the first
+/// render node. On a hybrid laptop it does not — the iGPU is bound first, so `renderD128` is
+/// Intel, and this importer opened a Mesa GBM display, asked it for a pbuffer-capable OpenGL
+/// config (Mesa's GBM platform only ever advertises window-capable ones), got none, and reported
+/// "no EGL config for OpenGL" on a machine whose NVIDIA EGL stack was working perfectly. Pick the
+/// device by what it *is*.
+///
+/// Deliberately a local scan rather than a `pf_gpu` dependency: this crate is a leaf (the import
+/// worker is its own process, spawned for fault isolation), and `pf_gpu::linux_render_node` answers
+/// a different question anyway — it follows the operator's VAAPI/GPU *preference*, which may well
+/// name the iGPU. The answer needed here is not "which GPU should we use" but "where is CUDA".
+fn nvidia_render_node() -> std::path::PathBuf {
+    use std::path::{Path, PathBuf};
+    if let Some(p) = std::env::var_os("PUNKTFUNK_ZEROCOPY_RENDER_NODE").filter(|s| !s.is_empty()) {
+        return PathBuf::from(p);
+    }
+    // No NVIDIA node found — sysfs may simply not be mounted (a container without /sys). Keep the
+    // historical guess; the CUDA context below is what actually fails if this host has no NVIDIA.
+    nvidia_render_node_in(Path::new("/dev/dri"), Path::new("/sys/class/drm"))
+        .unwrap_or_else(|| PathBuf::from("/dev/dri/renderD128"))
+}
+
+/// The scan half of [`nvidia_render_node`], parameterized on its two roots so a test can pin the
+/// sysfs shape it depends on (`<sys_class_drm>/renderD129/device/vendor` holding `0x10de`). Nodes
+/// are visited in name order so the choice is stable across boots.
+fn nvidia_render_node_in(
+    dri: &std::path::Path,
+    sys_class_drm: &std::path::Path,
+) -> Option<std::path::PathBuf> {
+    let mut nodes: Vec<std::ffi::OsString> = std::fs::read_dir(dri)
+        .map(|rd| {
+            rd.filter_map(|e| e.ok())
+                .map(|e| e.file_name())
+                .filter(|n| n.as_encoded_bytes().starts_with(b"renderD"))
+                .collect()
+        })
+        .unwrap_or_default();
+    nodes.sort();
+    nodes
+        .into_iter()
+        .find(|node| {
+            std::fs::read_to_string(sys_class_drm.join(node).join("device").join("vendor"))
+                .ok()
+                .and_then(|s| u32::from_str_radix(s.trim().trim_start_matches("0x"), 16).ok())
+                == Some(PCI_VENDOR_NVIDIA)
+        })
+        .map(|node| dri.join(node))
+}
+
 /// Per-size GL machinery to blit a dmabuf EGLImage into a CUDA-registrable `GL_RGBA8` texture.
 struct GlBlit {
    program: u32,
@@ -528,13 +584,15 @@ impl EglImporter {
        // GBM platform on the NVIDIA render node: this ties the EGLDisplay (and its GL contexts)
        // to the same DRM device CUDA-GL interop associates with, which the EGL device platform
        // did not (cuGraphicsGLRegisterImage rejected device-platform GL textures).
-        let path = std::ffi::CString::new("/dev/dri/renderD128").unwrap();
-        // SAFETY: `path` is a live local `CString` (built from a string with no interior NUL, so it
+        let node = nvidia_render_node();
+        let path = std::ffi::CString::new(node.as_os_str().as_encoded_bytes())
+            .with_context(|| format!("render node path {} has an interior NUL", node.display()))?;
+        // SAFETY: `path` is a live local `CString` (its constructor rejected interior NULs, so it
        // is NUL-terminated); `path.as_ptr()` is a valid pointer to that buffer which outlives this
        // synchronous `open`. `open` only reads the path and returns a new fd (or -1); it neither
        // retains the pointer nor writes through it, so there is no aliasing or lifetime hazard.
        let render_fd = unsafe { libc::open(path.as_ptr(), libc::O_RDWR | libc::O_CLOEXEC) };
-        ensure!(render_fd >= 0, "open /dev/dri/renderD128 for GBM");
+        ensure!(render_fd >= 0, "open {} for GBM", node.display());
        // SAFETY: `render_fd` is the live DRM render-node fd just returned by `open` and checked
        // `>= 0`. `gbm_create_device` (libgbm, linked above) builds a `gbm_device` over that fd and
        // returns a `*mut gbm_device` (or null); it borrows but does not take ownership of the fd,
@@ -546,7 +604,7 @@ impl EglImporter {
            // and no `EglImporter` exists yet to close it again, so this `close` runs exactly once on
            // the live `render_fd`, releasing it before the error return. No double-close.
            unsafe { libc::close(render_fd) };
-            anyhow::bail!("gbm_create_device failed");
+            anyhow::bail!("gbm_create_device failed on {}", node.display());
        }

        // SAFETY: `Egl::load_required` dlopens the system libEGL and binds its entry points,
@@ -568,7 +626,7 @@ impl EglImporter {
                &[egl::ATTRIB_NONE],
            )
        }
-        .context("eglGetPlatformDisplay(GBM) on the NVIDIA render node")?;
+        .with_context(|| format!("eglGetPlatformDisplay(GBM) on {}", node.display()))?;
        egl.initialize(display).context("eglInitialize")?;

        let exts = egl
@@ -590,20 +648,46 @@ impl EglImporter {
        egl.bind_api(egl::OPENGL_API)
            .context("eglBindAPI(OpenGL)")?;
        // The default EGL_SURFACE_TYPE in eglChooseConfig is WINDOW_BIT, which a headless device
-        // display has none of — request a pbuffer-capable config (we run surfaceless anyway).
-        let config = egl
-            .choose_first_config(
-                display,
-                &[
-                    egl::SURFACE_TYPE,
-                    egl::PBUFFER_BIT,
-                    egl::RENDERABLE_TYPE,
-                    egl::OPENGL_BIT,
-                    egl::NONE,
-                ],
-            )
+        // display has none of — ask for a pbuffer-capable config first (NVIDIA's GBM platform has
+        // them, and this is the request every working host has been served for a year).
+        //
+        // Then fall back to no surface-type constraint at all, because a config's surface type is
+        // irrelevant to us: we never create an EGLSurface, we `eglMakeCurrent` surfaceless. Mesa's
+        // GBM platform advertises ONLY window-capable configs (gbm_surface is the whole point of
+        // the platform there), so the pbuffer request finds nothing on any Mesa device — which is
+        // how a hybrid host that landed on the iGPU reported "no EGL config for OpenGL" while its
+        // NVIDIA EGL stack was fine. The node is picked correctly now; this keeps the failure from
+        // being fatal (and unreadable) if some other driver makes the same choice Mesa did.
+        let want_pbuffer = [
+            egl::SURFACE_TYPE,
+            egl::PBUFFER_BIT,
+            egl::RENDERABLE_TYPE,
+            egl::OPENGL_BIT,
+            egl::NONE,
+        ];
+        let any_surface = [egl::RENDERABLE_TYPE, egl::OPENGL_BIT, egl::NONE];
+        let config = match egl
+            .choose_first_config(display, &want_pbuffer)
            .context("eglChooseConfig")?
-            .context("no EGL config for OpenGL")?;
+        {
+            Some(c) => c,
+            None => {
+                tracing::debug!(
+                    node = %node.display(),
+                    "no pbuffer-capable OpenGL EGL config — retrying without a surface-type \
+                     constraint (we run surfaceless)"
+                );
+                egl.choose_first_config(display, &any_surface)
+                    .context("eglChooseConfig (no surface-type constraint)")?
+                    .with_context(|| {
+                        format!(
+                            "no EGL config for OpenGL on {} — this display serves no \
+                             OpenGL-renderable config at all",
+                            node.display()
+                        )
+                    })?
+            }
+        };
        let gl_ctx = egl
            .create_context(
                display,
@@ -637,6 +721,7 @@ impl EglImporter {
        // `eglCreateImage(EGL_LINUX_DMA_BUF_EXT)` requires as its context argument later.
        let no_ctx = unsafe { egl::Context::from_ptr(egl::NO_CONTEXT) };
        tracing::info!(
+            node = %node.display(),
            "zero-copy EGL importer ready (GBM platform + GL texture interop, dma_buf_import + modifiers)"
        );
        Ok(EglImporter {
@@ -1087,3 +1172,69 @@ impl Drop for EglImporter {
        }
    }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::path::{Path, PathBuf};
+
+    /// A fake `/dev/dri` + `/sys/class/drm` pair: every `(node, vendor)` gets a device node file
+    /// and, when `vendor` is `Some`, the sysfs `device/vendor` that identifies it.
+    fn fixture(nodes: &[(&str, Option<&str>)]) -> (tempfile::TempDir, PathBuf, PathBuf) {
+        let tmp = tempfile::tempdir().unwrap();
+        let dri = tmp.path().join("dev/dri");
+        let sys = tmp.path().join("sys/class/drm");
+        std::fs::create_dir_all(&dri).unwrap();
+        for (node, vendor) in nodes {
+            std::fs::write(dri.join(node), b"").unwrap();
+            if let Some(v) = vendor {
+                let dev = sys.join(node).join("device");
+                std::fs::create_dir_all(&dev).unwrap();
+                std::fs::write(dev.join("vendor"), v).unwrap();
+            }
+        }
+        (tmp, dri, sys)
+    }
+
+    /// The hybrid-laptop case this selection exists for: the iGPU is bound first, so the NVIDIA
+    /// GPU is NOT renderD128 — which is what the old hardcoded path assumed.
+    #[test]
+    fn picks_the_nvidia_node_not_the_first_one() {
+        let (_t, dri, sys) = fixture(&[
+            ("renderD128", Some("0x8086\n")),
+            ("renderD129", Some("0x10de\n")),
+        ]);
+        assert_eq!(
+            nvidia_render_node_in(&dri, &sys),
+            Some(dri.join("renderD129"))
+        );
+    }
+
+    /// No NVIDIA GPU (or no readable sysfs) ⇒ no answer, and the caller keeps the historical
+    /// `/dev/dri/renderD128` guess rather than inventing one.
+    #[test]
+    fn no_nvidia_node_yields_nothing() {
+        let (_t, dri, sys) = fixture(&[("renderD128", Some("0x8086\n")), ("renderD129", None)]);
+        assert_eq!(nvidia_render_node_in(&dri, &sys), None);
+        // Missing roots entirely (a container without /sys, or without /dev/dri).
+        assert_eq!(
+            nvidia_render_node_in(Path::new("/nonexistent/dri"), &sys),
+            None
+        );
+    }
+
+    /// Ignore card/control nodes, and visit render nodes in name order so a two-NVIDIA host picks
+    /// the same one every boot.
+    #[test]
+    fn scans_render_nodes_only_and_in_order() {
+        let (_t, dri, sys) = fixture(&[
+            ("card0", Some("0x10de\n")),
+            ("renderD130", Some("0x10de\n")),
+            ("renderD129", Some("0x10de\n")),
+        ]);
+        assert_eq!(
+            nvidia_render_node_in(&dri, &sys),
+            Some(dri.join("renderD129"))
+        );
+    }
+}
@@ -232,6 +232,48 @@ pub fn gpu_import_disabled() -> bool {
    GPU_IMPORT_DISABLED.load(Ordering::Relaxed)
 }

+/// The same idea for the OTHER zero-copy half: consecutive failures to import a raw dmabuf in the
+/// encoder (VAAPI's libva import, PyroWave's Vulkan one) with no successful frame in between.
+///
+/// That import can fail for reasons no retry can fix — a driver that will not take what the
+/// compositor allocates. The encode-stall recovery above it cannot tell the difference, so it
+/// rebuilt the identical failing encoder five times and then ended the video session, on every
+/// connection, forever: a hybrid Intel/NVIDIA laptop reporting `vaCreateSurfaces` →
+/// `VA_STATUS_ERROR_ALLOCATION_FAILED` on its first frame could not stream at all until its
+/// operator found `PUNKTFUNK_ZEROCOPY=0` by hand. The host already knows how to encode that
+/// machine — capture just has to stop handing it dmabufs. Latching here is what makes the next
+/// session negotiate CPU frames on its own.
+static RAW_DMABUF_FAILURE_STREAK: AtomicU32 = AtomicU32::new(0);
+static RAW_DMABUF_DISABLED: AtomicBool = AtomicBool::new(false);
+/// Below the encoder's own rebuild budget, so the latch is set before the session it doomed ends.
+const RAW_DMABUF_FAILURE_LATCH: u32 = 3;
+
+/// Record an encoder-side raw-dmabuf import failure. Latches the process-wide disable after
+/// [`RAW_DMABUF_FAILURE_LATCH`] consecutive failures.
+pub fn note_raw_dmabuf_import_failure(reason: &str) {
+    let streak = RAW_DMABUF_FAILURE_STREAK.fetch_add(1, Ordering::Relaxed) + 1;
+    if streak >= RAW_DMABUF_FAILURE_LATCH && !RAW_DMABUF_DISABLED.swap(true, Ordering::Relaxed) {
+        tracing::error!(
+            streak,
+            reason,
+            "zero-copy raw-dmabuf passthrough disabled for this host process: the encoder failed \
+             to import the compositor's dmabuf {streak} times in a row — captures fall back to the \
+             CPU path (slower, but this host could not stream at all otherwise)"
+        );
+    }
+}
+
+/// Record a raw dmabuf that imported and encoded — resets the failure streak.
+pub fn note_raw_dmabuf_import_ok() {
+    RAW_DMABUF_FAILURE_STREAK.store(0, Ordering::Relaxed);
+}
+
+/// True once repeated encoder import failures latched the raw-dmabuf passthrough off (see
+/// [`note_raw_dmabuf_import_failure`]).
+pub fn raw_dmabuf_import_disabled() -> bool {
+    RAW_DMABUF_DISABLED.load(Ordering::Relaxed)
+}
+
 /// DRM FourCC for a packed 32-bit format name (little-endian, e.g. `b"XR24"`).
 const fn fourcc(c: &[u8; 4]) -> u32 {
    (c[0] as u32) | ((c[1] as u32) << 8) | ((c[2] as u32) << 16) | ((c[3] as u32) << 24)
@@ -1813,9 +1813,11 @@ pub unsafe extern "C" fn punktfunk_generate_identity(
            return PunktfunkStatus::InvalidArg;
        }
        unsafe {
-            std::ptr::copy_nonoverlapping(cert.as_ptr(), cert_pem_out as *mut u8, cert.len());
+            // `.cast()`, not `as *mut u8`: `c_char` is i8 on x86_64 but u8 on aarch64, so the
+            // `as` form is a REQUIRED conversion on one and a no-op clippy rejects on the other.
+            std::ptr::copy_nonoverlapping(cert.as_ptr(), cert_pem_out.cast::<u8>(), cert.len());
            *cert_pem_out.add(cert.len()) = 0;
-            std::ptr::copy_nonoverlapping(key.as_ptr(), key_pem_out as *mut u8, key.len());
+            std::ptr::copy_nonoverlapping(key.as_ptr(), key_pem_out.cast::<u8>(), key.len());
            *key_pem_out.add(key.len()) = 0;
        }
        PunktfunkStatus::Ok
@@ -13,7 +13,13 @@
 //!   its rolling baseline: standing queue growth, the *pre-loss* signature of a saturated link
 //!   (bufferbloat) — this is the early-warning signal loss-based control lacks;
 //! - **a jump-to-live flush** — the pump discarded its backlog, the strongest "we were behind"
-//!   evidence there is.
+//!   evidence there is;
+//! - **host-encode-latency rise** — the host's per-AU 0xCF `encode_us` climbing above its rolling
+//!   baseline: the ENCODER falling behind its frame budget (the compute knee), the one failure a
+//!   fat LAN never surfaces as loss/OWD/decode. Paired with the host's own climb refusal (a
+//!   behind-cadence host acks climbs at the current rate) and short-ack cap learning
+//!   ([`BitrateController::on_ack`]), this is what stops an Automatic session from driving the
+//!   encoder off a cliff the network could carry.
 //!
 //! AIMD shape: a SEVERE window (an unrecoverable frame, a flush, ≥6 % loss, or a decode-latency
 //! excursion far past baseline) backs off ×0.7 immediately; ordinary congestion
@@ -94,6 +100,25 @@ const UTILIZATION_DEN: u64 = 4;
 /// cap always leaves ≥ ~12 % of climbing room — the two gates can't deadlock.
 const PROVEN_HEADROOM_NUM: u32 = 3;
 const PROVEN_HEADROOM_DEN: u32 = 2;
+/// How far the window's mean HOST-ENCODE latency (the 0xCF `HostStages::encode_us` the host
+/// already ships per AU) may rise above its rolling baseline before the window is bad. This is
+/// the down-driver for the ENCODER's compute knee — the failure loss/OWD/decode are all blind
+/// to: on a fat LAN the controller can climb to a rate the link carries fine but the ASIC
+/// can't encode inside the frame budget (4K120 HEVC at ~800 Mbps ≈ 9.3 ms against 8.33), and
+/// the only symptom is encode time. Baseline-RELATIVE on purpose: an escalated host reports
+/// encode_us inflated by its retrieve-queue depth (~a frame), so an absolute budget threshold
+/// would read permanently-red and drive the rate to the floor; a rise above the session's own
+/// baseline survives that offset. ~half a 120 Hz frame budget of standing rise is real.
+const ENCODE_RISE_US: i64 = 4_000;
+/// Host-encode latency this far above baseline (≈1.5 × a 120 Hz budget) is SEVERE — the encode
+/// queue is growing past the knee; skip the two-window confirmation.
+const ENCODE_SEVERE_US: i64 = 12_000;
+/// Clean windows parked at the learned [`host cap`](BitrateController::host_cap_kbps) before
+/// re-probing above it (~60 s at the 750 ms tick). A cadence-refusal cap is scene-dependent
+/// evidence, not a spec limit — without a re-probe, one heavy scene would cap the whole
+/// session. A still-standing limit just re-teaches itself in two short acks, which the host
+/// pre-clamps without touching the encoder — the re-probe costs no rebuild, no IDR.
+const CAP_REPROBE_WINDOWS: u32 = 80;
 /// Rolling window (in 750 ms report windows, ~30 s) whose minimum mean is the OWD baseline.
 /// Long enough to remember the uncongested floor, short enough to follow genuine path changes.
 const BASELINE_WINDOWS: usize = 40;
@@ -121,6 +146,27 @@ pub(crate) struct BitrateController {
    /// keeping-up baseline. Empty on embedders that don't report decode latency (the decode
    /// signal is then simply absent — identical to the pre-decode-signal behavior).
    decode_means: VecDeque<i64>,
+    /// Recent window mean host-encode latencies (µs, from the 0xCF datagrams); rolling-min
+    /// baseline like the decode signal. Cleared whenever OUR OWN rate decrease changes the
+    /// encode regime (see [`on_ack`](Self::on_ack)) and on a mode switch.
+    encode_means: VecDeque<i64>,
+    /// The host-taught rate cap (§ABR overdrive): latched when the host acks BELOW what we
+    /// asked twice consecutively at the same value — its encoder's codec-level ceiling, or a
+    /// climb refusal while host encode can't hold cadence. Kept apart from `ceiling_kbps` so
+    /// the probe-measured link authority survives a mode switch's reset. Slowly re-probed
+    /// ([`CAP_REPROBE_WINDOWS`]) so scene-dependent evidence can't cap the session forever.
+    host_cap_kbps: Option<u32>,
+    /// The rate the last [`request`](Self::request) asked for — the reference an ack is judged
+    /// short against. Taken (not kept) by the ack, so one request is judged at most once.
+    last_requested_kbps: Option<u32>,
+    /// Consecutive short-ack streak: the value and how many times in a row it was acked. Two
+    /// identical short acks latch [`host_cap_kbps`](Self::host_cap_kbps) — one can be a
+    /// transient (a failed host rebuild keeping the old rate); the host's resolves are
+    /// deterministic min()s, so a persistent limit reproduces exactly.
+    short_ack_kbps: u32,
+    short_acks: u32,
+    /// Clean windows spent parked at the learned cap (the re-probe clock).
+    cap_probe_windows: u32,
    /// Proven throughput: the session's highest windowed ACTUAL delivered rate seen with flat
    /// decode latency — the known-good high-water mark climbs are bounded against. Never decays;
    /// shrinking capacity (thermals, a heavier scene) is the reactive decode signal's job. On
@@ -147,6 +193,12 @@ impl BitrateController {
            probing: true,
            owd_means: VecDeque::with_capacity(BASELINE_WINDOWS),
            decode_means: VecDeque::with_capacity(BASELINE_WINDOWS),
+            encode_means: VecDeque::with_capacity(BASELINE_WINDOWS),
+            host_cap_kbps: None,
+            last_requested_kbps: None,
+            short_ack_kbps: 0,
+            short_acks: 0,
+            cap_probe_windows: 0,
            proven_kbps: 0,
            bad_windows: 0,
            clean_windows: 0,
@@ -168,22 +220,68 @@ impl BitrateController {

    /// The host's [`crate::quic::BitrateChanged`] ack: its clamp is authoritative for what the
    /// encoder now targets, and any ack proves the host renegotiates (resets the silence counter).
+    ///
+    /// A SHORT ack (below what we asked) is the host telling us about a limit the network
+    /// signals can't see — its encoder's codec-level ceiling, or a climb refusal while encode
+    /// can't hold cadence. Two consecutive short acks at the SAME value latch it as
+    /// [`host_cap_kbps`](Self::host_cap_kbps), stopping the AIMD sawtooth from re-poking a
+    /// limit the host already refused; ONE is not enough — a failed host rebuild also acks
+    /// short once, and latching a transient would cap the session on a hiccup.
    pub(crate) fn on_ack(&mut self, kbps: u32) {
        if kbps > 0 {
+            if kbps < self.current_kbps {
+                // Our own decrease changes the encode-time regime (less work per frame; on an
+                // escalated host the queue offset shifts too) — judging the new regime against
+                // the old baseline would train-fire the encode down-driver. Re-seed it.
+                self.encode_means.clear();
+            }
+            if let Some(req) = self.last_requested_kbps.take() {
+                if kbps < req {
+                    if self.short_ack_kbps == kbps {
+                        self.short_acks += 1;
+                    } else {
+                        self.short_ack_kbps = kbps;
+                        self.short_acks = 1;
+                    }
+                    if self.short_acks >= 2 && self.host_cap_kbps.is_none_or(|c| kbps < c) {
+                        tracing::info!(
+                            cap_kbps = kbps,
+                            "adaptive bitrate: host cap learned (encoder ceiling or cadence \
+                             refusal) — climbs stop here until it lifts"
+                        );
+                        self.host_cap_kbps = Some(kbps.max(self.floor_kbps));
+                        self.cap_probe_windows = 0;
+                    }
+                } else {
+                    self.short_acks = 0;
+                }
+            }
            self.current_kbps = kbps;
        }
        self.unacked = 0;
    }

+    /// An accepted mode switch: the encoder's ceiling and compute knee are properties of the
+    /// MODE (4K120 caps where 1080p60 never would) — drop the mode-scoped learned state. The
+    /// probe-measured `ceiling_kbps` (a LINK property) survives.
+    pub(crate) fn on_mode_switch(&mut self) {
+        self.host_cap_kbps = None;
+        self.short_acks = 0;
+        self.cap_probe_windows = 0;
+        self.encode_means.clear();
+    }
+
    /// Feed one report window; returns the rate to request now, if any. `dropped` = frames that
    /// went FEC-unrecoverable in the window, `loss_ppm` the window's [`crate::quic::LossReport`]
    /// figure, `owd_mean_us` the window's mean skew-corrected capture→received latency (`None`
    /// without a clock handshake), `decode_mean_us` the window's mean client decode-stage latency
-    /// (`None` on an embedder that doesn't report it — the signal is then absent), `actual_kbps`
-    /// the window's ACTUAL delivered throughput (wire bytes received ÷ window — what the pipeline
-    /// really carried, as opposed to the target it was allowed; feeds the utilization climb gate
-    /// and the proven-throughput high-water mark), `flushed` = the pump's jump-to-live fired in
-    /// the window.
+    /// (`None` on an embedder that doesn't report it — the signal is then absent),
+    /// `encode_mean_us` the window's mean HOST encode-stage latency (from the per-AU 0xCF
+    /// datagrams; `None` on an old host that doesn't send them), `actual_kbps` the window's
+    /// ACTUAL delivered throughput (wire bytes received ÷ window — what the pipeline really
+    /// carried, as opposed to the target it was allowed; feeds the utilization climb gate and
+    /// the proven-throughput high-water mark), `flushed` = the pump's jump-to-live fired in the
+    /// window.
    #[allow(clippy::too_many_arguments)]
    pub(crate) fn on_window(
        &mut self,
@@ -192,6 +290,7 @@ impl BitrateController {
        loss_ppm: u32,
        owd_mean_us: Option<i64>,
        decode_mean_us: Option<i64>,
+        encode_mean_us: Option<i64>,
        actual_kbps: u32,
        flushed: bool,
    ) -> Option<u32> {
@@ -242,6 +341,23 @@ impl BitrateController {
            }
            None => (false, false),
        };
+        // Host-encode latency: the same rolling-min-baseline treatment, measuring the HOST'S
+        // encoder — the compute-knee down-driver (see [`ENCODE_RISE_US`]). This is the only
+        // signal that can push an already-too-high rate back under the knee: the host refuses
+        // further climbs while behind cadence, but nothing else ever DESCENDS on a clean LAN.
+        let (encode_bad, encode_severe) = match encode_mean_us {
+            Some(mean) => {
+                let base = self.encode_means.iter().min().copied();
+                let bad = base.is_some_and(|b| mean > b + ENCODE_RISE_US);
+                let severe = base.is_some_and(|b| mean > b + ENCODE_SEVERE_US);
+                if self.encode_means.len() == BASELINE_WINDOWS {
+                    self.encode_means.pop_front();
+                }
+                self.encode_means.push_back(mean);
+                (bad, severe)
+            }
+            None => (false, false),
+        };
        // The proven-throughput high-water mark: this window's delivered rate is now demonstrably
        // digestible (decode latency stayed flat while it was carried). Loss doesn't disqualify —
        // the bytes that DID arrive still went through the decoder; what loss means for the rate
@@ -253,8 +369,9 @@ impl BitrateController {
        // deep decode-latency excursion) or loss far past any blip — one window is enough.
        // Ordinary congestion (heavy-but-recoverable loss, an OWD rise, a decode-latency rise)
        // still needs two consecutive windows.
-        let severe = dropped > 0 || flushed || loss_ppm >= SEVERE_LOSS_PPM || decode_severe;
-        let bad = severe || loss_ppm >= HEAVY_LOSS_PPM || owd_bad || decode_bad;
+        let severe =
+            dropped > 0 || flushed || loss_ppm >= SEVERE_LOSS_PPM || decode_severe || encode_severe;
+        let bad = severe || loss_ppm >= HEAVY_LOSS_PPM || owd_bad || decode_bad || encode_bad;
        if bad {
            self.bad_windows += 1;
            self.clean_windows = 0;
@@ -264,6 +381,29 @@ impl BitrateController {
            self.clean_windows += 1;
            self.bad_windows = 0;
        }
+        // The learned host cap re-probe (see [`CAP_REPROBE_WINDOWS`]): after ~60 s of clean
+        // windows parked at the cap, lift it one step (+12.5 %, ceiling-bounded) so a
+        // scene-dependent refusal can't quietly cap the whole session — a still-standing limit
+        // just re-latches from the next pair of short acks, at zero encoder cost.
+        if let Some(cap) = self.host_cap_kbps {
+            if bad {
+                self.cap_probe_windows = 0;
+            } else if self.current_kbps >= cap.saturating_sub(cap / 16) {
+                self.cap_probe_windows += 1;
+                if self.cap_probe_windows >= CAP_REPROBE_WINDOWS {
+                    self.cap_probe_windows = 0;
+                    let lifted = cap.saturating_add(cap / 8).min(self.ceiling_kbps);
+                    if lifted > cap {
+                        tracing::debug!(
+                            from_kbps = cap,
+                            to_kbps = lifted,
+                            "adaptive bitrate: re-probing above the learned host cap"
+                        );
+                        self.host_cap_kbps = Some(lifted);
+                    }
+                }
+            }
+        }
        let cooled = self
            .last_change
            .is_none_or(|t| now.duration_since(t) >= CHANGE_COOLDOWN);
@@ -284,10 +424,14 @@ impl BitrateController {
        // utilized window after a long-enough clean run climbs immediately.
        let utilized =
            actual_kbps as u64 * UTILIZATION_DEN >= self.current_kbps as u64 * UTILIZATION_NUM;
-        let cap = self
+        // The effective ceiling folds in the host-taught cap: the probe measured the LINK, but
+        // the host's short acks measured the ENCODER — whichever binds first is the limit.
+        let eff_ceiling = self
            .ceiling_kbps
+            .min(self.host_cap_kbps.unwrap_or(u32::MAX));
+        let cap = eff_ceiling
            .min(self.proven_kbps.saturating_mul(PROVEN_HEADROOM_NUM) / PROVEN_HEADROOM_DEN);
-        if self.current_kbps < self.ceiling_kbps && utilized && cap > self.current_kbps {
+        if self.current_kbps < eff_ceiling && utilized && cap > self.current_kbps {
            // Slow start: double on every cooled clean window until the first congestion signal
            // (this is how an Automatic session reaches a probe-measured ceiling in seconds).
            // Congestion avoidance: +~6 % after a sustained clean run.
@@ -308,6 +452,7 @@ impl BitrateController {
    fn request(&mut self, kbps: u32, now: Instant) -> Option<u32> {
        self.last_change = Some(now);
        self.unacked += 1;
+        self.last_requested_kbps = Some(kbps);
        // `current_kbps` is NOT updated here — the host's ack is authoritative. A lost/ignored
        // request just recomputes from the same base next time (and counts toward MAX_UNACKED).
        Some(kbps)
@@ -331,7 +476,16 @@ mod tests {
    fn run_clean(c: &mut BitrateController, start: Instant, from: u32, n: u32) -> Option<u32> {
        let mut out = None;
        for i in from..from + n {
-            out = c.on_window(ticks(start, i), 0, 0, Some(10_000), None, 1_000_000, false);
+            out = c.on_window(
+                ticks(start, i),
+                0,
+                0,
+                Some(10_000),
+                None,
+                None,
+                1_000_000,
+                false,
+            );
            if out.is_some() {
                return out;
            }
@@ -345,7 +499,7 @@ mod tests {
        let mut c = BitrateController::new(0);
        let now = Instant::now();
        assert_eq!(
-            c.on_window(now, 5, 900_000, Some(500_000), None, 1_000_000, true),
+            c.on_window(now, 5, 900_000, Some(500_000), None, None, 1_000_000, true),
            None
        );
    }
@@ -356,22 +510,58 @@ mod tests {
        let start = Instant::now();
        // Heavy-but-recoverable loss (2–6 %) is ORDINARY: one window is a blip — no reaction.
        assert_eq!(
-            c.on_window(ticks(start, 0), 0, 25_000, None, None, 1_000_000, false),
+            c.on_window(
+                ticks(start, 0),
+                0,
+                25_000,
+                None,
+                None,
+                None,
+                1_000_000,
+                false
+            ),
            None
        );
        // The second consecutive bad window backs off ×0.7.
        assert_eq!(
-            c.on_window(ticks(start, 1), 0, 25_000, None, None, 1_000_000, false),
+            c.on_window(
+                ticks(start, 1),
+                0,
+                25_000,
+                None,
+                None,
+                None,
+                1_000_000,
+                false
+            ),
            Some(14_000)
        );
        c.on_ack(14_000);
        // Still bad after the cooldown → another ×0.7 step from the ACKED rate.
        assert_eq!(
-            c.on_window(ticks(start, 6), 0, 25_000, None, None, 1_000_000, false),
+            c.on_window(
+                ticks(start, 6),
+                0,
+                25_000,
+                None,
+                None,
+                None,
+                1_000_000,
+                false
+            ),
            None
        ); // bad #1 again
        assert_eq!(
-            c.on_window(ticks(start, 7), 0, 25_000, None, None, 1_000_000, false),
+            c.on_window(
+                ticks(start, 7),
+                0,
+                25_000,
+                None,
+                None,
+                None,
+                1_000_000,
+                false
+            ),
            Some(9_800)
        );
    }
@@ -382,19 +572,28 @@ mod tests {
        let mut c = BitrateController::new(20_000);
        let start = Instant::now();
        assert_eq!(
-            c.on_window(ticks(start, 0), 1, 0, None, None, 1_000_000, false),
+            c.on_window(ticks(start, 0), 1, 0, None, None, None, 1_000_000, false),
            Some(14_000)
        );
        // …and so does a jump-to-live flush.
        let mut c = BitrateController::new(20_000);
        assert_eq!(
-            c.on_window(ticks(start, 0), 0, 0, None, None, 1_000_000, true),
+            c.on_window(ticks(start, 0), 0, 0, None, None, None, 1_000_000, true),
            Some(14_000)
        );
        // …and ≥6 % window loss.
        let mut c = BitrateController::new(20_000);
        assert_eq!(
-            c.on_window(ticks(start, 0), 0, 80_000, None, None, 1_000_000, false),
+            c.on_window(
+                ticks(start, 0),
+                0,
+                80_000,
+                None,
+                None,
+                None,
+                1_000_000,
+                false
+            ),
            Some(14_000)
        );
    }
@@ -404,18 +603,18 @@ mod tests {
        let mut c = BitrateController::new(20_000);
        let start = Instant::now();
        assert_eq!(
-            c.on_window(ticks(start, 0), 1, 0, None, None, 1_000_000, false),
+            c.on_window(ticks(start, 0), 1, 0, None, None, None, 1_000_000, false),
            Some(14_000)
        );
        c.on_ack(14_000);
        // A severe window INSIDE the 1.5 s cooldown (tick 1 = 750 ms) → held; at the cooldown
        // boundary (tick 2 = 1.5 s) it fires.
        assert_eq!(
-            c.on_window(ticks(start, 1), 1, 0, None, None, 1_000_000, false),
+            c.on_window(ticks(start, 1), 1, 0, None, None, None, 1_000_000, false),
            None
        );
        assert_eq!(
-            c.on_window(ticks(start, 2), 1, 0, None, None, 1_000_000, false),
+            c.on_window(ticks(start, 2), 1, 0, None, None, None, 1_000_000, false),
            Some(9_800)
        );
    }
@@ -426,17 +625,17 @@ mod tests {
        let start = Instant::now();
        // ×0.7 of 6000 = 4200 < floor → clamped to 5000.
        assert_eq!(
-            c.on_window(ticks(start, 0), 1, 0, None, None, 1_000_000, false),
+            c.on_window(ticks(start, 0), 1, 0, None, None, None, 1_000_000, false),
            Some(5_000)
        );
        c.on_ack(5_000);
        // At the floor, further bad windows request nothing.
        assert_eq!(
-            c.on_window(ticks(start, 6), 1, 0, None, None, 1_000_000, false),
+            c.on_window(ticks(start, 6), 1, 0, None, None, None, 1_000_000, false),
            None
        );
        assert_eq!(
-            c.on_window(ticks(start, 7), 1, 0, None, None, 1_000_000, false),
+            c.on_window(ticks(start, 7), 1, 0, None, None, None, 1_000_000, false),
            None
        );
    }
@@ -446,7 +645,7 @@ mod tests {
        let mut c = BitrateController::new(20_000);
        let start = Instant::now();
        assert_eq!(
-            c.on_window(ticks(start, 0), 1, 0, None, None, 1_000_000, false),
+            c.on_window(ticks(start, 0), 1, 0, None, None, None, 1_000_000, false),
            Some(14_000)
        );
        c.on_ack(14_000);
@@ -469,9 +668,16 @@ mod tests {
        // Every cooled clean window doubles until the ceiling caps the climb, then quiet.
        let mut got = Vec::new();
        for i in 0..14 {
-            if let Some(k) =
-                c.on_window(ticks(start, i), 0, 0, Some(10_000), None, 1_000_000, false)
-            {
+            if let Some(k) = c.on_window(
+                ticks(start, i),
+                0,
+                0,
+                Some(10_000),
+                None,
+                None,
+                1_000_000,
+                false,
+            ) {
                c.on_ack(k);
                got.push(k);
            }
@@ -485,20 +691,47 @@ mod tests {
        c.set_ceiling(300_000);
        let start = Instant::now();
        assert_eq!(
-            c.on_window(ticks(start, 0), 0, 0, Some(10_000), None, 1_000_000, false),
+            c.on_window(
+                ticks(start, 0),
+                0,
+                0,
+                Some(10_000),
+                None,
+                None,
+                1_000_000,
+                false
+            ),
            Some(40_000)
        );
        c.on_ack(40_000);
        // Severe window → immediate ×0.7, and slow start is over.
        assert_eq!(
-            c.on_window(ticks(start, 2), 1, 0, Some(10_000), None, 1_000_000, false),
+            c.on_window(
+                ticks(start, 2),
+                1,
+                0,
+                Some(10_000),
+                None,
+                None,
+                1_000_000,
+                false
+            ),
            Some(28_000)
        );
        c.on_ack(28_000);
        // Clean again — but the next climb is additive, after the 6-window clean run.
        let mut next = None;
        for i in 3..12 {
-            next = c.on_window(ticks(start, i), 0, 0, Some(10_000), None, 1_000_000, false);
+            next = c.on_window(
+                ticks(start, i),
+                0,
+                0,
+                Some(10_000),
+                None,
+                None,
+                1_000_000,
+                false,
+            );
            if next.is_some() {
                assert!(i >= 8, "additive climb must wait for the clean run");
                break;
@@ -512,7 +745,7 @@ mod tests {
        let mut c = BitrateController::new(0);
        c.set_ceiling(1_000_000);
        assert_eq!(
-            c.on_window(Instant::now(), 0, 0, None, None, 1_000_000, false),
+            c.on_window(Instant::now(), 0, 0, None, None, None, 1_000_000, false),
            None
        );
        let mut c = BitrateController::new(20_000);
@@ -527,17 +760,44 @@ mod tests {
        // Establish a ~10 ms baseline over a few clean windows.
        for i in 0..4 {
            assert_eq!(
-                c.on_window(ticks(start, i), 0, 0, Some(10_000), None, 1_000_000, false),
+                c.on_window(
+                    ticks(start, i),
+                    0,
+                    0,
+                    Some(10_000),
+                    None,
+                    None,
+                    1_000_000,
+                    false
+                ),
                None
            );
        }
        // Delay climbs 40 ms above baseline with ZERO loss — bufferbloat. Two windows → back off.
        assert_eq!(
-            c.on_window(ticks(start, 4), 0, 0, Some(50_000), None, 1_000_000, false),
+            c.on_window(
+                ticks(start, 4),
+                0,
+                0,
+                Some(50_000),
+                None,
+                None,
+                1_000_000,
+                false
+            ),
            None
        );
        assert_eq!(
-            c.on_window(ticks(start, 5), 0, 0, Some(52_000), None, 1_000_000, false),
+            c.on_window(
+                ticks(start, 5),
+                0,
+                0,
+                Some(52_000),
+                None,
+                None,
+                1_000_000,
+                false
+            ),
            Some(14_000)
        );
    }
@@ -557,6 +817,7 @@ mod tests {
                    0,
                    Some(10_000),
                    Some(8_000),
+                    None,
                    1_000_000,
                    false
                ),
@@ -572,6 +833,7 @@ mod tests {
                0,
                Some(10_000),
                Some(38_000),
+                None,
                1_000_000,
                false
            ),
@@ -584,6 +846,7 @@ mod tests {
                0,
                Some(10_000),
                Some(40_000),
+                None,
                1_000_000,
                false
            ),
@@ -605,6 +868,7 @@ mod tests {
                0,
                Some(10_000),
                Some(8_000),
+                None,
                1_000_000,
                false
            ),
@@ -620,6 +884,7 @@ mod tests {
                0,
                Some(10_000),
                Some(38_000),
+                None,
                1_000_000,
                false
            ),
@@ -634,6 +899,7 @@ mod tests {
                0,
                Some(10_000),
                Some(40_000),
+                None,
                1_000_000,
                false
            ),
@@ -657,6 +923,7 @@ mod tests {
                    0,
                    Some(10_000),
                    Some(8_000),
+                    None,
                    2_000,
                    false
                ),
@@ -673,6 +940,7 @@ mod tests {
                0,
                Some(10_000),
                Some(8_000),
+                None,
                18_000,
                false
            ),
@@ -696,6 +964,7 @@ mod tests {
                0,
                Some(10_000),
                Some(8_000),
+                None,
                20_000,
                false
            ),
@@ -710,6 +979,7 @@ mod tests {
                0,
                Some(10_000),
                Some(8_000),
+                None,
                30_000,
                false
            ),
@@ -732,6 +1002,7 @@ mod tests {
                0,
                Some(10_000),
                Some(8_000),
+                None,
                20_000,
                false
            ),
@@ -741,7 +1012,16 @@ mod tests {
        // A long calm stretch (2 % utilization, decoder idle): the controller stays silent.
        for i in 2..30 {
            assert_eq!(
-                c.on_window(ticks(start, i), 0, 0, Some(10_000), Some(4_000), 600, false),
+                c.on_window(
+                    ticks(start, i),
+                    0,
+                    0,
+                    Some(10_000),
+                    Some(4_000),
+                    None,
+                    600,
+                    false
+                ),
                None
            );
        }
@@ -761,6 +1041,7 @@ mod tests {
                    0,
                    Some(10_000),
                    Some(8_000),
+                    None,
                    1_000_000,
                    false
                ),
@@ -776,6 +1057,7 @@ mod tests {
                0,
                Some(10_000),
                Some(60_000),
+                None,
                1_000_000,
                false
            ),
@@ -783,6 +1065,235 @@ mod tests {
        );
    }

+    #[test]
+    fn two_identical_short_acks_latch_the_host_cap() {
+        // The 4K120 field failure: the encoder ceilings at ~794 Mbps while the link carries
+        // more — the host acks short. TWO identical short acks teach the cap; climbs then stop
+        // poking a limit the host already refused (the rebuild-storm driver).
+        let mut c = BitrateController::new(400_000);
+        c.set_ceiling(1_400_000);
+        let start = Instant::now();
+        assert_eq!(run_clean(&mut c, start, 0, 1), Some(800_000));
+        // First short ack: current follows (authoritative), but one short ack is not a cap.
+        c.on_ack(794_000);
+        assert!(c.host_cap_kbps.is_none());
+        // The next climb overshoots again and is short-acked at the SAME value: latch.
+        assert_eq!(run_clean(&mut c, start, 10, 1), Some(1_400_000));
+        c.on_ack(794_000);
+        assert_eq!(c.host_cap_kbps, Some(794_000));
+        // Parked AT the learned cap, nothing left to climb to — no more requests.
+        assert_eq!(run_clean(&mut c, start, 20, 12), None);
+    }
+
+    #[test]
+    fn one_short_ack_is_a_transient_not_a_cap() {
+        // A failed host rebuild acks short once (it kept the old rate) — latching THAT would
+        // cap the session on a driver hiccup. The streak must survive only identical repeats.
+        let mut c = BitrateController::new(400_000);
+        c.set_ceiling(1_400_000);
+        let start = Instant::now();
+        assert_eq!(run_clean(&mut c, start, 0, 1), Some(800_000));
+        c.on_ack(400_000); // rebuild failed, host kept the old rate
+        assert!(c.host_cap_kbps.is_none());
+        // The retry applies fully: streak broken, still no cap, full authority kept.
+        assert_eq!(run_clean(&mut c, start, 10, 1), Some(800_000));
+        c.on_ack(800_000);
+        assert!(c.host_cap_kbps.is_none());
+    }
+
+    #[test]
+    fn mode_switch_clears_the_learned_cap() {
+        let mut c = BitrateController::new(400_000);
+        c.set_ceiling(1_400_000);
+        let start = Instant::now();
+        assert_eq!(run_clean(&mut c, start, 0, 1), Some(800_000));
+        c.on_ack(794_000);
+        assert_eq!(run_clean(&mut c, start, 10, 1), Some(1_400_000));
+        c.on_ack(794_000);
+        assert_eq!(c.host_cap_kbps, Some(794_000));
+        // 4K120's ceiling means nothing at the new mode — the cap must not survive the switch
+        // (the probe-measured link ceiling does).
+        c.on_mode_switch();
+        assert!(c.host_cap_kbps.is_none());
+        assert_eq!(c.ceiling_kbps, 1_400_000);
+    }
+
+    #[test]
+    fn learned_cap_reprobes_after_a_sustained_clean_run() {
+        // A cadence-refusal cap is scene evidence, not a spec limit: after ~60 s parked clean
+        // at the cap, lift one step so a one-time heavy scene can't cap the session forever. A
+        // still-standing limit just re-latches from the next short-ack pair, at zero cost.
+        let mut c = BitrateController::new(400_000);
+        c.set_ceiling(1_400_000);
+        let start = Instant::now();
+        assert_eq!(run_clean(&mut c, start, 0, 1), Some(800_000));
+        c.on_ack(794_000);
+        assert_eq!(run_clean(&mut c, start, 10, 1), Some(1_400_000));
+        c.on_ack(794_000);
+        assert_eq!(c.host_cap_kbps, Some(794_000));
+        for i in 0..CAP_REPROBE_WINDOWS {
+            let _ = c.on_window(
+                ticks(start, 20 + i),
+                0,
+                0,
+                Some(10_000),
+                None,
+                None,
+                1_000_000,
+                false,
+            );
+        }
+        assert_eq!(c.host_cap_kbps, Some(794_000 + 794_000 / 8));
+    }
+
+    #[test]
+    fn host_encode_latency_rise_backs_off() {
+        // The compute knee: link pristine, client decoder fine — only HOST encode time moves
+        // (the 4K120 case: ~9.3 ms against an 8.33 ms budget shows up nowhere else). Two risen
+        // windows → ×0.7, exactly like an OWD/decode rise.
+        let mut c = BitrateController::new(20_000);
+        let start = Instant::now();
+        for i in 0..4 {
+            assert_eq!(
+                c.on_window(
+                    ticks(start, i),
+                    0,
+                    0,
+                    Some(10_000),
+                    None,
+                    Some(7_000),
+                    1_000_000,
+                    false
+                ),
+                None
+            );
+        }
+        assert_eq!(
+            c.on_window(
+                ticks(start, 4),
+                0,
+                0,
+                Some(10_000),
+                None,
+                Some(11_500),
+                1_000_000,
+                false
+            ),
+            None
+        );
+        assert_eq!(
+            c.on_window(
+                ticks(start, 6),
+                0,
+                0,
+                Some(10_000),
+                None,
+                Some(12_000),
+                1_000_000,
+                false
+            ),
+            Some(14_000)
+        );
+    }
+
+    #[test]
+    fn deep_encode_excursion_is_severe() {
+        // Encode time shooting ≈1.5 frame budgets over baseline = the queue is growing past
+        // the knee right now — no two-window confirmation.
+        let mut c = BitrateController::new(20_000);
+        let start = Instant::now();
+        for i in 0..4 {
+            assert_eq!(
+                c.on_window(
+                    ticks(start, i),
+                    0,
+                    0,
+                    Some(10_000),
+                    None,
+                    Some(7_000),
+                    1_000_000,
+                    false
+                ),
+                None
+            );
+        }
+        assert_eq!(
+            c.on_window(
+                ticks(start, 4),
+                0,
+                0,
+                Some(10_000),
+                None,
+                Some(20_000),
+                1_000_000,
+                false
+            ),
+            Some(14_000)
+        );
+    }
+
+    #[test]
+    fn rate_decrease_rebases_the_encode_baseline() {
+        // After OUR OWN decrease the encode regime legitimately changes (less work per frame;
+        // an escalated host's reported encode_us also carries a queue offset) — the old
+        // baseline must not train-fire repeated backoffs down to the floor.
+        let mut c = BitrateController::new(20_000);
+        let start = Instant::now();
+        for i in 0..4 {
+            let _ = c.on_window(
+                ticks(start, i),
+                0,
+                0,
+                Some(10_000),
+                None,
+                Some(7_000),
+                1_000_000,
+                false,
+            );
+        }
+        let _ = c.on_window(
+            ticks(start, 4),
+            0,
+            0,
+            Some(10_000),
+            None,
+            Some(12_000),
+            1_000_000,
+            false,
+        );
+        assert_eq!(
+            c.on_window(
+                ticks(start, 6),
+                0,
+                0,
+                Some(10_000),
+                None,
+                Some(12_500),
+                1_000_000,
+                false
+            ),
+            Some(14_000)
+        );
+        // The decrease applies → rebase. The new regime's ~15 ms means (an escalated host's
+        // queue offset) would be far over the OLD 7 ms baseline, but must now read clean.
+        c.on_ack(14_000);
+        for i in 8..11 {
+            assert_eq!(
+                c.on_window(
+                    ticks(start, i),
+                    0,
+                    0,
+                    Some(10_000),
+                    None,
+                    Some(15_000),
+                    1_000_000,
+                    false
+                ),
+                None
+            );
+        }
+    }
+
    #[test]
    fn ack_silence_disables_the_controller() {
        let mut c = BitrateController::new(20_000);
@@ -791,7 +1302,7 @@ mod tests {
        let mut i = 0;
        // Keep every window bad and never ack: exactly MAX_UNACKED requests, then silence.
        while i < 60 {
-            if c.on_window(ticks(start, i), 1, 0, None, None, 1_000_000, false)
+            if c.on_window(ticks(start, i), 1, 0, None, None, None, 1_000_000, false)
                .is_some()
            {
                sent += 1;
@@ -224,6 +224,20 @@ pub(crate) struct DecodeLatAcc {
    pub(crate) count: u32,
 }

+/// Host encode-stage latency accumulator — [`DecodeLatAcc`]'s mirror for the HOST side of the
+/// pipeline. The datagram task adds one sample per 0xCF `HostStages::encode_us` (host encoder
+/// submit → bitstream ready) and the pump drains a window mean into
+/// [`crate::abr::BitrateController::on_window`]'s encode signal. Host encode time was measured,
+/// shipped and drawn on the overlay, but never an ABR input — which is how a fat-LAN Automatic
+/// session drove the encoder past its compute knee with nothing to stop it (§ABR overdrive).
+/// Its own accumulator rather than the overlay's `host_timing` channel: that channel is a lossy
+/// `try_send` the embedder may never drain, and the controller must not depend on it.
+#[derive(Default)]
+pub(crate) struct EncodeLatAcc {
+    pub(crate) sum_us: u64,
+    pub(crate) count: u32,
+}
+
 /// The pre-decode video hand-off from the data-plane pump to the embedder. Unlike the side planes
 /// (self-contained samples that drop the newest on overflow), video AUs are reference-chained under the
 /// host's infinite GOP: dropping ANY frame mid-stream corrupts every dependent frame until the next
@@ -112,6 +112,12 @@ pub(super) async fn run_pump(args: WorkerArgs) {
    // Adaptive bitrate ack slot: the control task parks the latest BitrateChanged here; the
    // pump's controller drains it on its report tick (`take()` — an ack is consumed once).
    let bitrate_ack: Arc<Mutex<Option<u32>>> = Arc::new(Mutex::new(None));
+    // Host-encode-latency accumulator (the ABR encode signal, see [`EncodeLatAcc`]): the
+    // datagram task adds one sample per 0xCF; the pump drains a window mean per report tick.
+    let encode_lat = Arc::new(Mutex::new(super::frame_channel::EncodeLatAcc::default()));
+    // Bumped by the control task on every accepted mode switch (the `clock_gen` pattern): the
+    // pump resets the controller's mode-scoped learned state (host cap, encode baseline).
+    let mode_gen = Arc::new(AtomicU32::new(0));

    // Control task (see [`control_task`]): the handshake stream stays open for mid-stream
    // renegotiation, speed tests, clock re-sync, and clipboard metadata.
@@ -128,6 +134,7 @@ pub(super) async fn run_pump(args: WorkerArgs) {
            clock_gen: clock_gen.clone(),
            clip_event_tx: clip_event_tx.clone(),
            cursor_shape_tx,
+            mode_gen: mode_gen.clone(),
        }
        .run(),
    );
@@ -141,6 +148,7 @@ pub(super) async fn run_pump(args: WorkerArgs) {
        hidout_tx,
        hdr_meta_tx,
        host_timing_tx,
+        encode_lat.clone(),
        cursor_state_tx,
    ));

@@ -176,6 +184,8 @@ pub(super) async fn run_pump(args: WorkerArgs) {
        clock_offset,
        clock_gen,
        decode_lat,
+        encode_lat,
+        mode_gen,
        frames_dropped,
        fec_recovered,
        bitrate_ack,
@@ -24,6 +24,10 @@ pub(super) struct ControlTask {
    /// Host cursor shapes ([`CursorShape`], sent on pointer-bitmap change) → the embedder's
    /// shape plane ([`NativeClient::next_cursor_shape`]).
    pub(super) cursor_shape_tx: std::sync::mpsc::SyncSender<crate::quic::CursorShape>,
+    /// Bumped on every ACCEPTED mode switch (the `clock_gen` pattern): the pump watches it and
+    /// resets the bitrate controller's mode-scoped learned state — the encoder ceiling / compute
+    /// knee it was taught belong to the OLD mode.
+    pub(super) mode_gen: Arc<AtomicU32>,
 }

 impl ControlTask {
@@ -40,6 +44,7 @@ impl ControlTask {
            clock_gen,
            clip_event_tx,
            cursor_shape_tx,
+            mode_gen,
        } = self;
        // Mid-stream clock re-sync (see [`ClockResync`]): a batch runs every
        // CLOCK_RESYNC_INTERVAL and whenever the pump asks (CtrlRequest::ClockResync after
@@ -88,6 +93,7 @@ impl ControlTask {
                    if let Ok(ack) = Reconfigured::decode(&msg) {
                        if ack.accepted {
                            *mode_slot.lock().unwrap() = ack.mode;
+                            mode_gen.fetch_add(1, Ordering::Relaxed);
                            tracing::info!(mode = ?ack.mode, "host accepted mode switch");
                        } else {
                            tracing::warn!(active = ?ack.mode, "host rejected mode switch");
@@ -19,6 +19,12 @@ pub(super) struct DataPump {
    pub(super) clock_offset: Arc<std::sync::atomic::AtomicI64>,
    pub(super) clock_gen: Arc<AtomicU32>,
    pub(super) decode_lat: Arc<Mutex<DecodeLatAcc>>,
+    /// Host encode-stage latency window accumulator (the ABR encode signal — see
+    /// [`super::super::frame_channel::EncodeLatAcc`]); fed by the datagram task.
+    pub(super) encode_lat: Arc<Mutex<super::super::frame_channel::EncodeLatAcc>>,
+    /// Accepted-mode-switch generation (control task bumps): a change resets the controller's
+    /// mode-scoped learned state ([`BitrateController::on_mode_switch`]).
+    pub(super) mode_gen: Arc<AtomicU32>,
    pub(super) frames_dropped: Arc<std::sync::atomic::AtomicU64>,
    pub(super) fec_recovered: Arc<std::sync::atomic::AtomicU64>,
    pub(super) bitrate_ack: Arc<Mutex<Option<u32>>>,
@@ -41,6 +47,8 @@ impl DataPump {
            clock_offset: pump_clock_offset,
            clock_gen: pump_clock_gen,
            decode_lat: pump_decode_lat,
+            encode_lat: pump_encode_lat,
+            mode_gen: pump_mode_gen,
            frames_dropped,
            fec_recovered,
            bitrate_ack,
@@ -127,6 +135,7 @@ impl DataPump {
        let mut clock_detector_armed = true;
        let mut resync_wanted = false;
        let mut seen_clock_gen = pump_clock_gen.load(Ordering::Relaxed);
+        let mut seen_mode_gen = pump_mode_gen.load(Ordering::Relaxed);
        // Standing-latency bleed (see StandingLatency): the third detector, for the small,
        // constant, loss-free OWD elevation the two jump-to-live detectors deliberately
        // tolerate (< QUEUE_HIGH frames, < FLUSH_LATENCY behind) — a sub-frame standing
@@ -338,9 +347,15 @@ impl DataPump {
                        );
                    }
                }
-                // Adaptive bitrate: drain any host ack first (its clamp is authoritative), then
-                // feed the controller this window's congestion signals; a decision becomes a
-                // SetBitrate on the control stream.
+                // Adaptive bitrate: an accepted mode switch first (it invalidates the
+                // mode-scoped learned state), then drain any host ack (its clamp is
+                // authoritative), then feed the controller this window's congestion signals; a
+                // decision becomes a SetBitrate on the control stream.
+                let mg = pump_mode_gen.load(Ordering::Relaxed);
+                if mg != seen_mode_gen {
+                    seen_mode_gen = mg;
+                    abr.on_mode_switch();
+                }
                if let Some(acked) = bitrate_ack.lock().unwrap().take() {
                    abr.on_ack(acked);
                }
@@ -356,6 +371,14 @@ impl DataPump {
                    *acc = DecodeLatAcc::default();
                    (count > 0).then(|| (sum / count as u64) as i64)
                };
+                // Same drain for the host-encode window (0xCF `encode_us` via the datagram
+                // task) — `None` on an old host that doesn't send stage timings.
+                let encode_mean_us = {
+                    let mut acc = pump_encode_lat.lock().unwrap();
+                    let (sum, count) = (acc.sum_us, acc.count);
+                    *acc = Default::default();
+                    (count > 0).then(|| (sum / count as u64) as i64)
+                };
                // The window's ACTUAL delivered throughput — what the pipeline really carried, vs
                // the target it was allowed. Wire bytes (headers + FEC) slightly overstate the
                // media rate the decoder ingests; acceptable for the climb gate / proven-mark
@@ -369,17 +392,19 @@ impl DataPump {
                    loss_ppm,
                    owd_mean_us,
                    decode_mean_us,
+                    encode_mean_us,
                    actual_kbps,
                    flush_in_window,
                ) {
                    // Log the window's signals alongside the decision so an on-glass session can
-                    // tell a decode-driven re-target (the new signal — decode_mean_us elevated with
+                    // tell a decode-/encode-driven re-target (the new signals — elevated with
                    // loss/OWD flat) from a network-driven one.
                    tracing::info!(
                        kbps,
                        loss_ppm,
                        owd_mean_us = owd_mean_us.unwrap_or(-1),
                        decode_mean_us = decode_mean_us.unwrap_or(-1),
+                        encode_mean_us = encode_mean_us.unwrap_or(-1),
                        actual_kbps,
                        flushed = flush_in_window,
                        "adaptive bitrate: requesting encoder re-target"
@@ -14,6 +14,9 @@ pub(super) async fn run(
    hidout_tx: std::sync::mpsc::SyncSender<crate::quic::HidOutput>,
    hdr_meta_tx: std::sync::mpsc::SyncSender<crate::quic::HdrMeta>,
    host_timing_tx: std::sync::mpsc::SyncSender<crate::quic::HostTiming>,
+    // The ABR encode signal's accumulator (see [`EncodeLatAcc`]) — fed HERE, not off
+    // `host_timing_tx`: that channel is the overlay's, lossy and embedder-drained.
+    encode_lat: Arc<Mutex<super::super::frame_channel::EncodeLatAcc>>,
    cursor_state_tx: std::sync::mpsc::SyncSender<crate::quic::CursorState>,
 ) {
    // Per-pad reorder gate for v2 rumble envelopes (the seq analog of the host's gamepad-state
@@ -74,6 +77,11 @@ pub(super) async fn run(
            }
            Some(&crate::quic::HOST_TIMING_MAGIC) => {
                if let Some(t) = crate::quic::decode_host_timing_datagram(&d) {
+                    if let Some(s) = &t.stages {
+                        let mut acc = encode_lat.lock().unwrap();
+                        acc.sum_us += s.encode_us as u64;
+                        acc.count += 1;
+                    }
                    let _ = host_timing_tx.try_send(t);
                }
            }
@@ -70,6 +70,24 @@ fn apply_hdr(base: u32, hdr: bool) -> u32 {
 /// negotiates a codec the encoder can't open. NVENC and the GPU-less software path keep the
 /// Moonlight-validated static superset. HDR (Main10) is layered on by [`codec_mode_support`].
 fn base_codec_mode_support() -> u32 {
+    // A GPU-less host encodes H.264 and nothing else (openh264), so advertising the superset made
+    // Moonlight negotiate HEVC/AV1 and the session then died at encoder open with "the software
+    // encoder emits H.264 only". `pf_encode::Codec::host_wire_caps` — the native plane's twin of
+    // this function — has gated on exactly this since it was written; this one never did.
+    //
+    // Deliberately a local gate rather than delegating wholesale to `host_wire_caps()`: that would
+    // be the drift-proof shape, but on Windows it re-runs the DXGI adapter enumeration several
+    // times per `/serverinfo` GET (the probe helpers each sample it), and this endpoint is polled.
+    // The software case is a plain config read, so it costs nothing here. (Follow-up worth doing:
+    // the static `MaxLumaPixelsHEVC` in the XML above still advertises an HEVC limit even when the
+    // mask drops HEVC — harmless, since Moonlight gates on the mask, but it is a second and now
+    // inconsistent advertisement.)
+    if matches!(
+        pf_host_config::config().encoder_pref.as_str(),
+        "software" | "sw" | "openh264"
+    ) {
+        return super::SCM_H264;
+    }
    #[cfg(target_os = "linux")]
    if crate::encode::linux_zero_copy_is_vaapi() {
        if let Some(m) = probed_mask(crate::encode::vaapi_codec_support()) {
@@ -686,6 +686,13 @@ fn stream_body(
        true,
    )
    .context("open video encoder for stream")?;
+    // Tell the encoder how deep the capturer lets it pipeline. Without this an in-place backend
+    // (Windows direct-NVENC, which encodes the capturer's textures with no CopyResource) bounds
+    // itself by an env cap instead of the ring it is actually reading, and the capturer rotates a
+    // texture out from under a live encode — torn/mixed frames, never an error. The backend now
+    // also fails safe when nobody tells it, but pass the REAL depth: `idd_depth` is configurable
+    // and a deeper ring is free pipelining the fallback would forfeit.
+    enc.set_input_ring_depth(capturer.pipeline_depth().max(1));
    // FEC overhead percent (Sunshine default 20). Override with PUNKTFUNK_FEC_PCT (0 = data-only).
    let fec_pct: u8 = std::env::var("PUNKTFUNK_FEC_PCT")
        .ok()
@@ -841,6 +848,8 @@ fn stream_body(
                    true,                         // metadata-cursor capture — see the first open
                )
                .context("reopen encoder after rebuild")?;
+                // A rebuilt encoder starts unconfigured — same reason as the first open above.
+                enc.set_input_ring_depth(capturer.pipeline_depth().max(1));
                supports_rfi = enc.caps().supports_rfi;
                enc.request_keyframe();
                last_keyframe = Some(Instant::now());
@@ -150,11 +150,13 @@ pub(crate) struct StreamInfo {
 pub(crate) struct LocalSummary {
    /// Host version (mirrors `/health`).
    version: String,
-    /// True while the video stream thread is running.
+    /// True while video is streaming on EITHER plane: the GameStream media pipeline, or a live
+    /// native (punktfunk/1) session — the default plane, invisible in the GameStream flag alone.
    video_streaming: bool,
-    /// True while the audio stream thread is running.
+    /// True while audio is streaming on either plane (same rule as `video_streaming`).
    audio_streaming: bool,
-    /// The active launch session (set by Moonlight's `/launch`, cleared on cancel/stop).
+    /// The active session: GameStream's launch (Moonlight `/launch`) when present, else the first
+    /// live native session. `null` when nothing is streaming.
    session: Option<SessionInfo>,
    /// Number of pinned (paired) GameStream client certificates.
    paired_clients: u32,
@@ -391,26 +393,27 @@ pub(crate) async fn get_status(State(st): State<Arc<MgmtState>>) -> Json<Runtime
    )
 )]
 pub(crate) async fn get_local_summary(State(st): State<Arc<MgmtState>>) -> Json<LocalSummary> {
+    // Native punktfunk/1 plane (the DEFAULT plane; GameStream is opt-in) — read ONCE and used for
+    // both the session card and the streaming flags below.
+    let native = crate::session_status::snapshot();
    // GameStream launch, else the first live native session — so the tray reflects a native session
    // too (same GameStream-only blind spot the Dashboard `/status` had; see `session_status`).
    let session = st
        .app
        .launch
        .lock()
-        .unwrap()
+        .unwrap_or_else(|e| e.into_inner())
        .map(|l| SessionInfo {
            width: l.width,
            height: l.height,
            fps: l.fps,
        })
        .or_else(|| {
-            crate::session_status::snapshot()
-                .first()
-                .map(|s| SessionInfo {
-                    width: s.width,
-                    height: s.height,
-                    fps: s.fps,
-                })
+            native.first().map(|s| SessionInfo {
+                width: s.width,
+                height: s.height,
+                fps: s.fps,
+            })
        });
    let (native_paired_clients, pending_approvals) = st
        .native
@@ -419,8 +422,10 @@ pub(crate) async fn get_local_summary(State(st): State<Arc<MgmtState>>) -> Json<
        .unwrap_or((0, 0));
    Json(LocalSummary {
        version: env!("PUNKTFUNK_VERSION").into(),
-        video_streaming: st.app.streaming.load(Ordering::SeqCst),
-        audio_streaming: st.app.audio_streaming.load(Ordering::SeqCst),
+        // Either plane counts, like `/status`: reading only the GameStream flags made the tray say
+        // "idle" (and wear the idle icon) through an entire native session.
+        video_streaming: st.app.streaming.load(Ordering::SeqCst) || !native.is_empty(),
+        audio_streaming: st.app.audio_streaming.load(Ordering::SeqCst) || !native.is_empty(),
        session,
        paired_clients: st
            .app
@@ -286,11 +286,74 @@ async fn health_is_open_and_versioned() {
    assert_eq!(body["abi_version"], punktfunk_core::ABI_VERSION);
 }

+/// Serializes the tests that read (or write) the process-global live-session registry
+/// ([`crate::session_status`]): a session registered by one test would otherwise make a
+/// concurrently running one see a stream it never started.
+static SESSION_REGISTRY_LOCK: tokio::sync::Mutex<()> = tokio::sync::Mutex::const_new(());
+
+/// A `/local/summary` request from a loopback peer (the tray's own).
+fn summary_req() -> axum::http::Request<Body> {
+    let mut req = get_req("/api/v1/local/summary");
+    req.extensions_mut()
+        .insert(PeerAddr("127.0.0.1:40000".parse().unwrap()));
+    req
+}
+
+/// Registers a stand-in live native session; the returned guard removes it on drop.
+fn fake_native_session(
+    width: u32,
+    height: u32,
+    fps: u32,
+) -> crate::session_status::LiveSessionGuard {
+    let packed = ((width as u64) << 32) | ((height as u64) << 16) | fps as u64;
+    crate::session_status::register(
+        Arc::new(std::sync::atomic::AtomicU64::new(packed)),
+        Arc::new(std::sync::atomic::AtomicU32::new(20_000)),
+        Codec::H265,
+        Arc::new(std::sync::atomic::AtomicBool::new(false)),
+        Arc::new(std::sync::atomic::AtomicBool::new(false)),
+        "test-client".into(),
+        false,
+        Arc::new(std::sync::atomic::AtomicU32::new(0)),
+        Arc::new(std::sync::atomic::AtomicU32::new(0)),
+    )
+}
+
+/// A native (punktfunk/1) session — the DEFAULT plane — must read as streaming in the tray's
+/// summary. The GameStream `streaming` flag stays false throughout such a session, and reading it
+/// alone left the tray showing "idle" (with the idle icon) for the whole stream: exactly the blind
+/// spot `/status` was fixed for in [`crate::session_status`], which `/local/summary` still had.
+#[tokio::test]
+async fn local_summary_reports_a_native_session_as_streaming() {
+    let _serial = SESSION_REGISTRY_LOCK.lock().await;
+    let app = test_app(test_state(), None);
+
+    let (status, body) = send(&app, summary_req()).await;
+    assert_eq!(status, StatusCode::OK);
+    assert_eq!(body["video_streaming"], false);
+    assert_eq!(body["session"], serde_json::Value::Null);
+
+    let session = fake_native_session(3840, 2160, 120);
+    let (_, body) = send(&app, summary_req()).await;
+    assert_eq!(body["video_streaming"], true, "native session: {body}");
+    assert_eq!(body["audio_streaming"], true, "native session: {body}");
+    assert_eq!(body["session"]["width"], 3840);
+    assert_eq!(body["session"]["height"], 2160);
+    assert_eq!(body["session"]["fps"], 120);
+
+    // Session over → back to idle.
+    drop(session);
+    let (_, body) = send(&app, summary_req()).await;
+    assert_eq!(body["video_streaming"], false);
+    assert_eq!(body["session"], serde_json::Value::Null);
+}
+
 /// The tray's `/local/summary` is unauthenticated for LOOPBACK peers only — a LAN peer is
 /// rejected even though the route needs no bearer token, and the body never carries secret
 /// material (no PIN values, no fingerprints, no device names — counts/booleans only).
 #[tokio::test]
 async fn local_summary_is_loopback_only_and_non_sensitive() {
+    let _serial = SESSION_REGISTRY_LOCK.lock().await;
    let np = Arc::new(
        crate::native_pairing::NativePairing::load_with(
            Some(std::env::temp_dir().join(format!("pf-mgmt-summary-{}.json", std::process::id()))),
@@ -600,6 +663,7 @@ async fn compositors_lists_all_backends_with_flags() {

 #[tokio::test]
 async fn status_reflects_runtime_state() {
+    let _serial = SESSION_REGISTRY_LOCK.lock().await;
    let state = test_state();
    let app = test_app(state.clone(), None);

@@ -756,6 +820,8 @@ async fn stop_session_clears_runtime_state() {

 #[tokio::test]
 async fn idr_requires_an_active_stream() {
+    // A live native session (registered by a sibling test) is an active stream to this route.
+    let _serial = SESSION_REGISTRY_LOCK.lock().await;
    let state = test_state();
    let app = test_app(state.clone(), None);
    let post = || {
@@ -985,6 +985,18 @@ async fn serve_session(
    // full IDR when the encoder supports it (native-AMF LTR / Windows NVENC).
    let (rfi_tx, rfi_rx) = std::sync::mpsc::channel::<(u32, u32)>();
    let (bitrate_tx, bitrate_rx) = std::sync::mpsc::channel::<u32>();
+    // Encoder-truth bridge, data plane → control task (§ABR overdrive). The encode loop publishes
+    // here; the control task reads at `SetBitrate`-resolve time, so the ack the client's
+    // controller climbs from tracks what the encoder ACTUALLY does, not what was asked:
+    // - `live_bitrate`: the encoder's applied rate (kbps) — also the send pacer's/console's view.
+    // - `encoder_ceiling_kbps`: the discovered codec-level ceiling (0 = none discovered yet);
+    //   resolves land at min(policy clamp, ceiling), so overshoots stop costing rebuilds.
+    // - `cadence_degraded`: encode can't hold the frame cadence — a climb is refused (acked at
+    //   the current rate); the network isn't the bottleneck, more bits are anti-medicine.
+    // Plain atomics, not a channel: only the freshest value matters, and only at resolve time.
+    let live_bitrate = Arc::new(AtomicU32::new(welcome.bitrate_kbps));
+    let encoder_ceiling_kbps = Arc::new(AtomicU32::new(0));
+    let cadence_degraded = Arc::new(AtomicBool::new(false));
    let (probe_tx, probe_rx) = std::sync::mpsc::channel::<ProbeRequest>();
    let (probe_result_tx, probe_result_rx) = tokio::sync::mpsc::unbounded_channel::<ProbeResult>();
    // Mode-switch outcome, data plane → control task (same pattern as `probe_result_tx`): the accept
@@ -1036,6 +1048,9 @@ async fn serve_session(
        live_reconfig_ok,
        adaptive_fec,
        session_bitrate_kbps,
+        live_bitrate.clone(),
+        encoder_ceiling_kbps.clone(),
+        cadence_degraded.clone(),
        fec_target_ctl,
        reconfig_tx,
        keyframe_tx,
@@ -1429,6 +1444,9 @@ async fn serve_session(
                        bitrate_rx,
                        compositor,
                        bitrate_kbps,
+                        live_bitrate,
+                        encoder_ceiling_kbps,
+                        cadence_degraded,
                        bitrate_auto,
                        bit_depth,
                        chroma,
@@ -22,6 +22,13 @@ pub(super) async fn run(
    live_reconfig_ok: bool,
    adaptive_fec: bool,
    session_bitrate_kbps: u32,
+    // Encoder-truth bridge (data plane → here, §ABR overdrive): the encoder's live applied rate,
+    // its discovered codec-level ceiling (0 = unknown), and the "encode can't hold cadence"
+    // flag. Read at `SetBitrate`-resolve time so the ack — the base the client's controller
+    // climbs from — never promises a rate the encoder won't run at.
+    live_bitrate: Arc<AtomicU32>,
+    encoder_ceiling_kbps: Arc<AtomicU32>,
+    cadence_degraded: Arc<AtomicBool>,
    fec_target_ctl: Arc<AtomicU8>,
    reconfig_tx: std::sync::mpsc::Sender<punktfunk_core::Mode>,
    keyframe_tx: std::sync::mpsc::Sender<()>,
@@ -161,7 +168,31 @@ pub(super) async fn run(
                        );
                        session_bitrate_kbps
                    } else {
-                        resolve_bitrate_kbps(req.bitrate_kbps)
+                        let mut r = resolve_bitrate_kbps(req.bitrate_kbps);
+                        // Encoder truth (§ABR overdrive): the ack below is the base the
+                        // client's controller climbs from, so it must not promise past the
+                        // encoder's discovered codec-level ceiling — the pre-fix path acked
+                        // 1.01 Gbps while the ASIC ran 794 Mbps, and the controller climbed
+                        // from the phantom number forever (a ~0.6 s rebuild + IDR per step).
+                        let ceiling = encoder_ceiling_kbps.load(Ordering::Relaxed);
+                        if ceiling != 0 && r > ceiling {
+                            r = ceiling;
+                        }
+                        // Climb refusal while encode can't hold cadence: on a fat LAN no
+                        // network signal ever stops the climb, and past the compute knee more
+                        // bits only deepen the miss. Resolve a CLIMB to the current applied
+                        // rate (descents pass — they're the cure); the short ack teaches the
+                        // client controller its ceiling.
+                        let live = live_bitrate.load(Ordering::Relaxed);
+                        if cadence_degraded.load(Ordering::Relaxed) && live != 0 && r > live {
+                            tracing::info!(
+                                requested_kbps = req.bitrate_kbps,
+                                held_kbps = live,
+                                "bitrate climb refused — encode is behind cadence"
+                            );
+                            r = live;
+                        }
+                        r
                    };
                    tracing::debug!(
                        requested_kbps = req.bitrate_kbps,
@@ -909,6 +909,20 @@ pub(super) struct SessionContext {
    pub(super) compositor: crate::vdisplay::Compositor,
    /// Negotiated encoder bitrate (kbps).
    pub(super) bitrate_kbps: u32,
+    /// The encoder's live APPLIED rate (kbps) — shared with the send pacer, the web console, the
+    /// mgmt registry AND the control task (which acks climbs against it). The encode loop stores
+    /// `Encoder::applied_bitrate_bps` here after every apply, so everything downstream tracks
+    /// what the ASIC really targets, not what was requested (§ABR overdrive).
+    pub(super) live_bitrate: Arc<AtomicU32>,
+    /// The encoder's discovered codec-level bitrate ceiling (kbps; 0 = none discovered): written
+    /// when an apply comes back short, read by this loop (pre-clamp incoming requests — a
+    /// request already AT the ceiling then costs nothing) and by the control task (truthful
+    /// acks from the first post-discovery request).
+    pub(super) encoder_ceiling_kbps: Arc<AtomicU32>,
+    /// "Encode can't hold the frame cadence" (the escalation leaky bucket is elevated, or the
+    /// session escalated): while set, the control task refuses bitrate CLIMBS — the network
+    /// isn't the bottleneck, feeding the encoder more bits deepens the miss.
+    pub(super) cadence_degraded: Arc<AtomicBool>,
    /// The client asked for "Automatic" (`Hello::bitrate_kbps == 0`), so `bitrate_kbps` came from
    /// the host's codec-aware default. For PyroWave that default is the ~1.6 bpp operating point of
    /// the NEGOTIATED MODE (`resolve_bitrate_kbps_for`) — a mid-stream mode switch re-resolves it
@@ -1035,6 +1049,9 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
        bitrate_rx,
        compositor,
        mut bitrate_kbps,
+        live_bitrate,
+        encoder_ceiling_kbps,
+        cadence_degraded,
        bitrate_auto,
        bit_depth,
        // The resolved chroma is already captured in `plan` (above); ignore the duplicate here.
@@ -1254,9 +1271,9 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
    // The bounded channel applies backpressure (the encode thread blocks if the send falls behind,
    // so frames slow down rather than a dropped frame freezing the infinite-GOP stream).
    let (frame_tx, frame_rx) = std::sync::mpsc::sync_channel::<SendMsg>(3);
-    // Live encoder bitrate, shared with the send thread's stats sample: a mid-stream adaptive
-    // bitrate change (bitrate_rx below) updates it so the console shows the actual target.
-    let live_bitrate = Arc::new(AtomicU32::new(bitrate_kbps));
+    // `live_bitrate` (SessionContext) is shared with the send thread's stats sample AND the
+    // control task: a mid-stream adaptive bitrate change (bitrate_rx below) stores the
+    // encoder-APPLIED rate, so the console, pacer and climb-refusal acks all see the truth.
    // Live session mode, same pattern (H3): a mid-stream mode switch (reconfig below) updates it so
    // a stats capture armed after a resize registers the real mode. Seeded with the refresh the
    // initial build actually achieved (`interval_hz`), not the request — KWin may cap a virtual
@@ -1436,6 +1453,26 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
    const DEPTH_ESCALATE: u32 = 20;
    const DEPTH_BEHIND_CAP: u32 = 60;
    const DEPTH_WARMUP_FRAMES: u64 = 60;
+    // Half the escalate threshold: ~10 net behind-frames is already solid "the encoder, not the
+    // network, is the bottleneck" evidence — enough to flag `cadence_degraded` (the control task
+    // then refuses bitrate CLIMBS) well before the session pays a latency escalation for it.
+    const DEPTH_DEGRADE: u32 = 10;
+    // De-escalation (the escalate-and-hold v1's missing half): a sustained clean run at the
+    // escalated setting (~5 s at 120 fps, every frame on cadence) earns ONE attempt at winding
+    // back — reverse order of the escalation, pipelined retrieve first (its rebuild restores
+    // sub-frame streaming and the IO-stream binding), then capture depth back to 1. Each
+    // attempt costs the wind-back rebuild's IDR, so attempts are paced by an exponential
+    // backoff (1 → 5 → 25 min, capped) — a workload that genuinely needs the escalation
+    // converges to keeping it, but NEVER a permanent latch: a latch plus the ABR sawtooth
+    // pinned sessions at the floor with the escalation stuck.
+    const DEESCALATE_CLEAN_FRAMES: u32 = 600;
+    const DEESCALATE_BACKOFF_START: std::time::Duration = std::time::Duration::from_secs(60);
+    const DEESCALATE_BACKOFF_MAX: std::time::Duration = std::time::Duration::from_secs(25 * 60);
+    let mut pipelined_active = false;
+    let mut deescalating = false;
+    let mut ahead_run: u32 = 0;
+    let mut deescalate_not_before: Option<std::time::Instant> = None;
+    let mut deescalate_backoff = DEESCALATE_BACKOFF_START;
    while !stop.load(Ordering::SeqCst) && std::time::Instant::now() < deadline {
        // Mid-stream session switch (the box flipped Gaming↔Desktop): rebuild the WHOLE backend in
        // place — a different compositor at the SAME client mode — keeping the Session + send thread
@@ -1674,15 +1711,49 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
        while let Ok(k) = bitrate_rx.try_recv() {
            want_kbps = Some(k);
        }
+        // Known-ceiling pre-clamp (§ABR overdrive): once the encoder's codec-level ceiling is
+        // known, resolve an over-asking request HERE — a request that clamps to the rate we're
+        // already at then skips the whole apply, where the pre-fix path bounced every overshoot
+        // off the driver into a full rebuild + IDR (~0.6 s each, four in one logged minute).
+        // (The control task clamps its acks from the same atomic; this covers requests already
+        // in flight when the ceiling was discovered.)
+        if let Some(k) = want_kbps.as_mut() {
+            let ceiling = encoder_ceiling_kbps.load(Ordering::Relaxed);
+            if ceiling != 0 && *k > ceiling {
+                tracing::info!(
+                    requested_kbps = *k,
+                    ceiling_kbps = ceiling,
+                    "bitrate request clamped to the known encoder ceiling"
+                );
+                *k = ceiling;
+            }
+        }
        if let Some(new_kbps) = want_kbps.filter(|&k| k != bitrate_kbps) {
            if enc.reconfigure_bitrate(new_kbps as u64 * 1000) {
+                // Adopt the encoder's post-clamp truth, not the request: it feeds the send
+                // pacer, the console/mgmt view and the control task's acks, and a short apply
+                // teaches the ceiling used above.
+                let applied_kbps = enc
+                    .applied_bitrate_bps()
+                    .map(|b| (b / 1000) as u32)
+                    .filter(|&k| k > 0)
+                    .unwrap_or(new_kbps);
                tracing::info!(
                    from_kbps = bitrate_kbps,
-                    to_kbps = new_kbps,
+                    to_kbps = applied_kbps,
+                    requested_kbps = new_kbps,
                    "encoder bitrate reconfigured in place (adaptive bitrate — no IDR)"
                );
-                bitrate_kbps = new_kbps;
-                live_bitrate.store(new_kbps, Ordering::Relaxed);
+                if applied_kbps < new_kbps {
+                    encoder_ceiling_kbps.store(applied_kbps, Ordering::Relaxed);
+                }
+                if applied_kbps < bitrate_kbps {
+                    // Down-step: the behind-cadence backlog was scored against the old,
+                    // heavier rate — clean slate so it can't feed a false escalation.
+                    behind_score = 0;
+                }
+                bitrate_kbps = applied_kbps;
+                live_bitrate.store(applied_kbps, Ordering::Relaxed);
                // Same encoder, same stream: the in-flight AUs and the wire-index prediction
                // stay valid — no inflight forfeit, no IDR-cooldown anchor.
            } else {
@@ -1702,9 +1773,17 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
                    plan.cursor_blend,
                ) {
                    Ok(mut new_enc) => {
+                        // The fresh encoder may have clamped to its codec-level ceiling —
+                        // adopt (and record) ITS rate, not the request; see the in-place arm.
+                        let applied_kbps = new_enc
+                            .applied_bitrate_bps()
+                            .map(|b| (b / 1000) as u32)
+                            .filter(|&k| k > 0)
+                            .unwrap_or(new_kbps);
                        tracing::info!(
                            from_kbps = bitrate_kbps,
-                            to_kbps = new_kbps,
+                            to_kbps = applied_kbps,
+                            requested_kbps = new_kbps,
                            "encoder rebuilt at new bitrate (adaptive bitrate)"
                        );
                        if let Some(c) = plan.wire_chunk {
@@ -1714,8 +1793,11 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
                        // directly so an ABR rebuild re-establishes the bound immediately.)
                        new_enc.set_input_ring_depth(capturer.pipeline_depth().max(1));
                        enc = new_enc;
-                        bitrate_kbps = new_kbps;
-                        live_bitrate.store(new_kbps, Ordering::Relaxed);
+                        if applied_kbps < new_kbps {
+                            encoder_ceiling_kbps.store(applied_kbps, Ordering::Relaxed);
+                        }
+                        bitrate_kbps = applied_kbps;
+                        live_bitrate.store(applied_kbps, Ordering::Relaxed);
                        // The owed AUs died with the old encoder — same bookkeeping as a
                        // mode-switch rebuild; the fresh encoder opens on an IDR, so anchor the
                        // IDR cooldown too.
@@ -1723,6 +1805,12 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
                        last_au_at = std::time::Instant::now();
                        encoder_resets = 0;
                        last_forced_idr = Some(std::time::Instant::now());
+                        // The rebuild stall itself (~0.6 s ≈ 70 missed deadlines at 120 fps,
+                        // 3.5× the escalate threshold) must not feed the contention
+                        // escalation — clean slate + re-run the warmup before judging again.
+                        behind_score = 0;
+                        depth_frames = 0;
+                        ahead_run = 0;
                    }
                    Err(e) => {
                        tracing::warn!(error = %format!("{e:#}"), to_kbps = new_kbps,
@@ -2547,7 +2635,7 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
        // two-thread lock moves the encode wait off this loop so capture/submit keep cadence,
        // at ~one tick of AU latency. `enc.set_pipelined` may decline (unsupported backend or
        // an explicit PUNKTFUNK_NVENC_ASYNC=0); either way it is asked exactly once.
-        if idd_adaptive_enabled() && (cur_depth < max_depth || !pipeline_asked) {
+        if idd_adaptive_enabled() {
            depth_frames += 1;
            if depth_frames > DEPTH_WARMUP_FRAMES {
                let behind = std::time::Instant::now() >= next;
@@ -2556,28 +2644,89 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
                } else {
                    behind_score.saturating_sub(1)
                };
-                if behind_score >= DEPTH_ESCALATE {
+                let escalated = cur_depth > 1 || pipelined_active || deescalating;
+                // Export "encode can't hold cadence" for the control task's climb refusal.
+                // An escalated session stays flagged even with the bucket drained: its climb
+                // headroom is spent, and letting climbs resume would saw against the
+                // escalation and starve the de-escalation clean run below.
+                cadence_degraded.store(
+                    escalated || behind_score >= DEPTH_DEGRADE,
+                    Ordering::Relaxed,
+                );
+                if deescalating {
+                    // A requested wind-back completes at the encoder's drained safe point —
+                    // poll it (the call is a cheap latch check until then).
+                    if !enc.set_pipelined(false) {
+                        deescalating = false;
+                        pipelined_active = false;
+                        // Re-arm the ask: a future sustained overrun may escalate again (the
+                        // backoff below paces how soon another wind-back may follow it).
+                        pipeline_asked = false;
+                        tracing::info!(
+                            "encoder pipelined retrieve de-escalated — sync retrieve (and \
+                             sub-frame streaming, where armed) restored; re-monitoring cadence"
+                        );
+                        // The wind-back rebuild's own stall must not re-escalate on the spot.
+                        behind_score = 0;
+                        depth_frames = 0;
+                        ahead_run = 0;
+                    }
+                } else if behind_score >= DEPTH_ESCALATE
+                    && (cur_depth < max_depth || !pipeline_asked)
+                {
                    if cur_depth < max_depth {
                        cur_depth = max_depth;
                        tracing::info!(
                            depth = cur_depth,
                            "IDD pipeline depth escalated — encode can't hold cadence at depth-1 \
-                             (GPU contention); pipelining for the rest of the session (latency \
+                             (GPU contention); pipelining until cadence holds clean (latency \
                             trade for throughput)"
                        );
                    } else {
                        pipeline_asked = true;
-                        if enc.set_pipelined(true) {
+                        pipelined_active = enc.set_pipelined(true);
+                        if pipelined_active {
                            tracing::info!(
                                "encoder pipelined retrieve escalated — encode can't hold \
                                 cadence and the capturer has no depth to give; the encode wait \
-                                 moves off the loop for the rest of the session (latency trade \
+                                 moves off the loop until cadence holds clean (latency trade \
                                 for throughput)"
                            );
                        }
                    }
                    // Give the action time to take effect before judging again.
                    behind_score = 0;
+                    ahead_run = 0;
+                } else if escalated {
+                    // De-escalation: a sustained every-frame-on-cadence run at the escalated
+                    // setting is the evidence the contention passed (a lower ABR rate, the
+                    // game scene lightened) — wind back in reverse order, paced by the
+                    // exponential backoff (see the consts above).
+                    ahead_run = if behind { 0 } else { ahead_run + 1 };
+                    if ahead_run >= DEESCALATE_CLEAN_FRAMES
+                        && deescalate_not_before.is_none_or(|t| std::time::Instant::now() >= t)
+                    {
+                        ahead_run = 0;
+                        deescalate_not_before =
+                            Some(std::time::Instant::now() + deescalate_backoff);
+                        deescalate_backoff = (deescalate_backoff * 5).min(DEESCALATE_BACKOFF_MAX);
+                        if pipelined_active {
+                            tracing::info!(
+                                "cadence held clean while escalated — winding the pipelined \
+                                 retrieve back (latency recovery; costs one IDR)"
+                            );
+                            deescalating = true;
+                        } else if cur_depth > 1 {
+                            cur_depth = 1;
+                            tracing::info!(
+                                depth = cur_depth,
+                                "IDD pipeline depth de-escalated — cadence held clean at the \
+                                 escalated depth (latency recovery)"
+                            );
+                            behind_score = 0;
+                            depth_frames = 0;
+                        }
+                    }
                }
            }
        }
@@ -431,16 +431,69 @@ fn web_setup(args: &[String]) -> Result<()> {
    ) {
        eprintln!("warning: could not add the firewall rule for TCP 47992");
    }
-    // 5. wait briefly for the host's mgmt token, then start (restart-on-failure picks it up otherwise)
-    for _ in 0..30 {
-        if token_path.exists() {
-            break;
+    // 5. Wait for EVERY file the launcher needs, then start the task and VERIFY it came up.
+    //
+    //    `web-run.cmd` refuses to serve without the mgmt token AND the host identity cert, and the
+    //    host writes them in that order: the token during argument parsing (`main::parse_serve`,
+    //    milliseconds after `serve` starts), the cert only once `serve` reaches
+    //    `gamestream::cert::ServerIdentity::load_or_create` - behind a pure-Rust RSA-2048 keygen
+    //    whose prime search has a multi-second tail. Gating on the token alone therefore fired
+    //    `schtasks /run` while that keygen was still running: the launcher exited 1, and since the
+    //    task carries no trigger other than boot (and Task Scheduler's restart-on-failure does not
+    //    reliably fire for a plain non-zero exit code), a freshly installed console stayed down
+    //    until the next reboot. Gate on `cert.pem` - written LAST, after `key.pem` - so all three
+    //    files are on disk before the first start.
+    let cert_path = data_dir.join("cert.pem");
+    if !wait_for_files(&[token_path.as_path(), cert_path.as_path()], 90) {
+        eprintln!(
+            "warning: the host service has not written {} + {} yet - not starting the console now; \
+             it will start at the next boot (or run: schtasks /run /tn {WEB_TASK})",
+            token_path.display(),
+            cert_path.display()
+        );
+        println!("web console set up (https://<host-ip>:47992)");
+        return Ok(());
+    }
+    if start_web_task() {
+        println!("web console set up + started (https://<host-ip>:47992)");
+    } else {
+        // Never claim "started" when it isn't: the launcher's own diagnostics go to a task with no
+        // console, so this warning is the only trace an operator gets at install time.
+        eprintln!(
+            "warning: the {WEB_TASK} task did not bring up a listener on :47992 - check its Last Run \
+             Result in Task Scheduler; the console will be retried at the next boot"
+        );
+        println!("web console set up (https://<host-ip>:47992)");
+    }
+    Ok(())
+}
+
+/// Poll until every path exists, or `secs` elapse. Returns whether they all showed up.
+fn wait_for_files(paths: &[&Path], secs: u64) -> bool {
+    for _ in 0..secs {
+        if paths.iter().all(|p| p.exists()) {
+            return true;
        }
        std::thread::sleep(std::time::Duration::from_secs(1));
    }
-    run_quiet("schtasks", &["/run", "/tn", WEB_TASK]);
-    println!("web console set up + started (https://<host-ip>:47992)");
-    Ok(())
+    paths.iter().all(|p| p.exists())
+}
+
+/// `schtasks /run` + verify: the command reports only that a start was *attempted*, so poll for the
+/// console's own listener and retry a couple of times before giving up. Returns whether :47992 ended
+/// up listening.
+fn start_web_task() -> bool {
+    for _ in 0..3 {
+        run_quiet("schtasks", &["/run", "/tn", WEB_TASK]);
+        // Bun binds the port a second or two after launch on a warm box; allow for a cold one.
+        for _ in 0..10 {
+            std::thread::sleep(std::time::Duration::from_secs(1));
+            if !web_listener_pids().is_empty() {
+                return true;
+            }
+        }
+    }
+    false
 }

 /// Source: a non-empty `--password-file` (fresh install) > keep existing (upgrade) > random fallback.
@@ -505,30 +558,57 @@ fn random_password() -> String {
 /// ("LISTENING"/"ABHOEREN"/...) is never parsed.
 fn stop_web_console() {
    run_quiet("schtasks", &["/end", "/tn", WEB_TASK]);
-    for line in run_capture("netstat", &["-ano", "-p", "tcp"]).lines() {
-        let toks: Vec<&str> = line.split_whitespace().collect();
-        if toks.len() >= 5
-            && toks[0].eq_ignore_ascii_case("tcp")
-            && toks[1].ends_with(":47992")
-            && (toks[2] == "0.0.0.0:0" || toks[2] == "[::]:0")
-        {
-            let pid = toks[toks.len() - 1];
-            if !pid.is_empty() && pid.bytes().all(|b| b.is_ascii_digit()) {
-                run_quiet("taskkill", &["/PID", pid, "/F"]);
-            }
-        }
+    for pid in web_listener_pids() {
+        run_quiet("taskkill", &["/PID", &pid, "/F"]);
    }
    std::thread::sleep(std::time::Duration::from_secs(1));
 }

-/// Register the boot/SYSTEM/restart-on-failure task via a generated Task Scheduler XML (`schtasks /xml`,
-/// no COM). The XML declares UTF-16, so it's written UTF-16LE+BOM.
+/// PIDs owning a LISTEN socket on :47992. Also the console's liveness probe (`start_web_task`).
+fn web_listener_pids() -> Vec<String> {
+    run_capture("netstat", &["-ano", "-p", "tcp"])
+        .lines()
+        .filter_map(|line| {
+            let toks: Vec<&str> = line.split_whitespace().collect();
+            let listening = toks.len() >= 5
+                && toks[0].eq_ignore_ascii_case("tcp")
+                && toks[1].ends_with(":47992")
+                && (toks[2] == "0.0.0.0:0" || toks[2] == "[::]:0");
+            let pid = *toks.last()?;
+            (listening && !pid.is_empty() && pid.bytes().all(|b| b.is_ascii_digit()))
+                .then(|| pid.to_string())
+        })
+        .collect()
+}
+
+/// Register the boot+logon/SYSTEM/restart-on-failure task via a generated Task Scheduler XML
+/// (`schtasks /xml`, no COM). The XML declares UTF-16, so it's written UTF-16LE+BOM.
+///
+/// Two triggers, because boot alone left too many holes: a console that lost the install-time start
+/// (or a box where the host service was started only later) had to wait for a full reboot. Logon is
+/// free insurance - `MultipleInstancesPolicy=IgnoreNew` makes a redundant start a no-op. If a
+/// Task Scheduler build rejects the two-trigger XML we fall back to the boot-only form rather than
+/// fail registration outright (no task at all is strictly worse than a boot-only task).
 fn register_web_task(cmd: &Path) -> Result<()> {
+    let cmd_xml = xml_escape(&cmd.to_string_lossy());
+    let boot = "<BootTrigger><Enabled>true</Enabled></BootTrigger>";
+    let boot_and_logon = "<BootTrigger><Enabled>true</Enabled></BootTrigger>\
+                          <LogonTrigger><Enabled>true</Enabled></LogonTrigger>";
+    if try_register_web_task(&cmd_xml, boot_and_logon) || try_register_web_task(&cmd_xml, boot) {
+        println!("registered scheduled task {WEB_TASK} -> {}", cmd.display());
+        Ok(())
+    } else {
+        bail!("schtasks /create {WEB_TASK} failed")
+    }
+}
+
+/// One `schtasks /create /xml` attempt with the given `<Triggers>` body. Returns whether it took.
+fn try_register_web_task(cmd_xml: &str, triggers: &str) -> bool {
    let xml = format!(
        "<?xml version=\"1.0\" encoding=\"UTF-16\"?>\n\
 <Task version=\"1.2\" xmlns=\"http://schemas.microsoft.com/windows/2004/02/mit/task\">\n\
  <RegistrationInfo><Description>punktfunk web management console (Nitro SSR on bun, :47992)</Description></RegistrationInfo>\n\
-  <Triggers><BootTrigger><Enabled>true</Enabled></BootTrigger></Triggers>\n\
+  <Triggers>{triggers}</Triggers>\n\
  <Principals><Principal id=\"Author\"><UserId>S-1-5-18</UserId><RunLevel>HighestAvailable</RunLevel></Principal></Principals>\n\
  <Settings>\n\
    <MultipleInstancesPolicy>IgnoreNew</MultipleInstancesPolicy>\n\
@@ -538,12 +618,13 @@ fn register_web_task(cmd: &Path) -> Result<()> {
    <ExecutionTimeLimit>PT0S</ExecutionTimeLimit>\n\
    <RestartOnFailure><Interval>PT1M</Interval><Count>10</Count></RestartOnFailure>\n\
  </Settings>\n\
-  <Actions Context=\"Author\"><Exec><Command>{}</Command></Exec></Actions>\n\
-</Task>",
-        xml_escape(&cmd.to_string_lossy())
+  <Actions Context=\"Author\"><Exec><Command>{cmd_xml}</Command></Exec></Actions>\n\
+</Task>"
    );
    let xml_path = std::env::temp_dir().join("punktfunk-web-task.xml");
-    write_utf16le_bom(&xml_path, &xml)?;
+    if write_utf16le_bom(&xml_path, &xml).is_err() {
+        return false;
+    }
    let ok = run_quiet(
        "schtasks",
        &[
@@ -556,12 +637,7 @@ fn register_web_task(cmd: &Path) -> Result<()> {
        ],
    );
    let _ = std::fs::remove_file(&xml_path);
-    if ok {
-        println!("registered scheduled task {WEB_TASK} -> {}", cmd.display());
-        Ok(())
-    } else {
-        bail!("schtasks /create {WEB_TASK} failed")
-    }
+    ok
 }

 fn write_utf16le_bom(path: &Path, s: &str) -> Result<()> {
@@ -15,8 +15,8 @@ use crate::status::{self, Poller, TrayStatus};
 struct HostTray {
    status: TrayStatus,
    web_port: u16,
-    /// The console answered the poller's live loopback probe — the "Open web console" entry is
-    /// shown iff opening it would actually work (repo-run consoles included, stopped ones not).
+    /// The console answered the poller's live loopback probe — labels the (always present) "Open
+    /// web console" entry; it never hides it.
    web_console: bool,
    /// Filled right after `spawn` (the poller needs the tray handle first) — lets menu actions
    /// force an immediate re-poll instead of waiting out the cadence.
@@ -33,8 +33,10 @@ impl HostTray {
        }
    }

-    fn open_console(&self) {
-        let url = format!("https://127.0.0.1:{}", self.web_port);
+    /// Open the web console at `path` ("" = dashboard). Deep links land the operator on the page
+    /// the menu entry promised — the pairing queue, the virtual displays.
+    fn open_console(&self, path: &str) {
+        let url = format!("https://127.0.0.1:{}/{path}", self.web_port);
        let _ = std::process::Command::new("xdg-open").arg(url).spawn();
    }
 }
@@ -51,7 +53,7 @@ impl ksni::Tray for HostTray {
    fn status(&self) -> ksni::Status {
        match &self.status {
            TrayStatus::Error(_) => ksni::Status::NeedsAttention,
-            s if s.pairing_attention() || s.has_conflicts() => ksni::Status::NeedsAttention,
+            s if s.pairing_attention() => ksni::Status::NeedsAttention,
            _ => ksni::Status::Active,
        }
    }
@@ -107,17 +109,33 @@ impl ksni::Tray for HostTray {
            }
            .into(),
            MenuItem::Separator,
+            // Always present — it is the reason most people open this menu. When the loopback
+            // probe says the console isn't answering, the label says so rather than the entry
+            // disappearing (a menu missing its main action reads as a broken tray).
            StandardItem {
-                label: "Open web console".into(),
-                visible: self.web_console,
-                activate: Box::new(|t: &mut Self| t.open_console()),
+                label: if self.web_console {
+                    "Open web console".to_string()
+                } else {
+                    "Open web console (not responding)".to_string()
+                },
+                activate: Box::new(|t: &mut Self| t.open_console("")),
                ..Default::default()
            }
            .into(),
            StandardItem {
                label: "Approve pairing request…".into(),
-                visible: self.web_console && self.status.pairing_attention(),
-                activate: Box::new(|t: &mut Self| t.open_console()),
+                visible: self.status.pairing_attention(),
+                activate: Box::new(|t: &mut Self| t.open_console("pairing")),
+                ..Default::default()
+            }
+            .into(),
+            StandardItem {
+                label: match self.status.kept_displays() {
+                    1 => "Release kept display…".to_string(),
+                    n => format!("Release {n} kept displays…"),
+                },
+                visible: self.status.kept_displays() > 0,
+                activate: Box::new(|t: &mut Self| t.open_console("displays")),
                ..Default::default()
            }
            .into(),
@@ -37,11 +37,6 @@ pub struct Summary {
    /// host that doesn't send it deserializes as 0.
    #[serde(default)]
    pub kept_displays: u32,
-    /// Other Moonlight-compatible hosts (Sunshine/Apollo/…) the host detected on this machine at
-    /// startup — side-by-side use is unsupported. `#[serde(default)]` so an older host omitting it
-    /// deserializes as empty.
-    #[serde(default)]
-    pub conflicts: Vec<String>,
 }

 #[derive(Clone, Copy, Debug, PartialEq, serde::Deserialize)]
@@ -74,47 +69,43 @@ impl TrayStatus {
            TrayStatus::Starting => "punktfunk host — starting…".into(),
            TrayStatus::Degraded => "punktfunk host — running (status unavailable)".into(),
            TrayStatus::Error(e) => format!("punktfunk host — failed ({e})"),
-            TrayStatus::Running(s) => {
-                let base = match (&s.session, s.video_streaming) {
-                    (Some(sess), true) => format!(
-                        "punktfunk host {} — streaming {}×{}@{}",
-                        s.version, sess.width, sess.height, sess.fps
-                    ),
-                    (_, true) => format!("punktfunk host {} — streaming", s.version),
-                    // Idle, but surface a kept (lingering/pinned) display: it — and, under an
-                    // exclusive topology, your physical monitors — is being held. Release it from
-                    // the console.
-                    _ if s.kept_displays > 0 => format!(
-                        "punktfunk host {} — idle · {} display{} kept",
-                        s.version,
-                        s.kept_displays,
-                        if s.kept_displays == 1 { "" } else { "s" }
-                    ),
-                    _ => format!("punktfunk host {} — idle", s.version),
-                };
-                // A conflicting Moonlight host (Sunshine/Apollo/…) is the loudest thing to say —
-                // side-by-side use is unsupported, so lead the tooltip with it.
-                if s.conflicts.is_empty() {
-                    base
-                } else {
-                    format!("⚠ conflicting host: {} — {base}", s.conflicts.join(", "))
-                }
-            }
+            TrayStatus::Running(s) => match (&s.session, self.is_streaming()) {
+                (Some(sess), true) => format!(
+                    "punktfunk host {} — streaming {}×{}@{}",
+                    s.version, sess.width, sess.height, sess.fps
+                ),
+                (_, true) => format!("punktfunk host {} — streaming", s.version),
+                // Idle, but surface a kept (lingering/pinned) display: it — and, under an
+                // exclusive topology, your physical monitors — is being held. Release it from
+                // the console.
+                _ if s.kept_displays > 0 => format!(
+                    "punktfunk host {} — idle · {} display{} kept",
+                    s.version,
+                    s.kept_displays,
+                    if s.kept_displays == 1 { "" } else { "s" }
+                ),
+                _ => format!("punktfunk host {} — idle", s.version),
+            },
        }
    }

-    /// The host detected another Moonlight-compatible host (Sunshine/Apollo/…) on this machine —
-    /// unsupported side-by-side. Drives the Linux (ksni) backend's `NeedsAttention` state; the
-    /// Windows backend surfaces the same conflict through the tooltip `headline()` instead (it has
-    /// no distinct attention icon), so this accessor is unused there — allow it per-platform rather
-    /// than gate the shared API out.
-    #[cfg_attr(not(target_os = "linux"), allow(dead_code))]
-    pub fn has_conflicts(&self) -> bool {
-        matches!(self, TrayStatus::Running(s) if !s.conflicts.is_empty())
+    /// A client is streaming: the host's flag, OR a live session in the summary.
+    ///
+    /// The session is checked too because a host from before the `get_local_summary` fix raised
+    /// `video_streaming` from the GameStream plane only — through a whole session on the native
+    /// (default) plane it said false while still reporting that session's mode, which is what left
+    /// the tray sitting at "idle" mid-stream. This keeps a newer tray honest against such a host.
+    pub fn is_streaming(&self) -> bool {
+        matches!(self, TrayStatus::Running(s) if s.video_streaming || s.session.is_some())
    }

-    pub fn is_streaming(&self) -> bool {
-        matches!(self, TrayStatus::Running(s) if s.video_streaming)
+    /// Virtual displays held with no live session (lingering/pinned) — offered as a one-click
+    /// release in the menu, since holding one can also be keeping physical monitors dark.
+    pub fn kept_displays(&self) -> u32 {
+        match self {
+            TrayStatus::Running(s) => s.kept_displays,
+            _ => 0,
+        }
    }

    /// A pairing attempt is waiting on the operator (shown as an extra menu entry).
@@ -156,9 +147,11 @@ struct Shared {

 impl Poller {
    /// Spawn the poll thread; `on_change(status, console_up)` fires (from that thread) whenever
-    /// either changes. `console_up` is a live loopback probe of the web console on `web_port` —
-    /// ground truth for the "Open web console" menu entry (a layout sniff would miss consoles run
-    /// from a repo checkout, and shows a dead entry while an installed console is still starting).
+    /// either changes. `console_up` is a live loopback probe of the web console on `web_port`. It
+    /// annotates the "Open web console" entry ("not responding") rather than hiding it: the entry
+    /// is the tray's most-wanted action, and a menu that silently drops it — because the console
+    /// was still starting, or the probe timed out — is indistinguishable from a tray that never
+    /// had one.
    pub fn spawn(
        mgmt_addr: String,
        mgmt_port: u16,
@@ -203,6 +196,10 @@ fn poll_loop(
    // When the summary became unreachable while the service was running (grace anchor).
    // Runs for the process lifetime (the tray exits by process exit; nothing to unwind).
    let mut unreachable_since: Option<Instant> = None;
+    // Consecutive failed console probes. One miss is not "down": the console is a bun/Nitro SSR
+    // whose cold first render can outrun this agent's 2 s timeout, and a menu entry that changes
+    // its label every few seconds reads as broken.
+    let mut console_misses = 0u32;
    loop {
        let svc = probe_service();
        let summary = if svc == ServiceState::Running {
@@ -219,7 +216,13 @@ fn poll_loop(
        };
        let grace_expired = unreachable_since.is_some_and(|t| t.elapsed() >= START_GRACE);
        let status = map_status(&svc, summary, grace_expired);
-        let console_up = probe_console(&agent, &console_url);
+        let console_up = if probe_console(&agent, &console_url) {
+            console_misses = 0;
+            true
+        } else {
+            console_misses += 1;
+            console_misses < 2
+        };
        if last.as_ref() != Some(&(status.clone(), console_up)) {
            on_change(status.clone(), console_up);
            last = Some((status, console_up));
@@ -406,7 +409,6 @@ mod tests {
            pin_pending: false,
            pending_approvals: 0,
            kept_displays: 0,
-            conflicts: Vec::new(),
        }
    }

@@ -449,15 +451,20 @@ mod tests {
        }
    }

+    /// Conflicting-host detection is deliberately NOT a tray concern: "installed" is not
+    /// "running", and an always-on ⚠ over a merely-present Sunshine cried wolf. The host still
+    /// detects and reports it (startup log, `detect-conflicts`, the mgmt summary) — the tray just
+    /// ignores that field, which it must keep tolerating on the wire.
    #[test]
-    fn conflicts_drive_attention_and_lead_the_tooltip() {
-        let mut s = summary(false);
-        assert!(!TrayStatus::Running(s.clone()).has_conflicts());
-        s.conflicts = vec!["Sunshine (running)".into(), "Apollo".into()];
-        let st = TrayStatus::Running(s);
-        assert!(st.has_conflicts());
-        let head = st.headline();
-        assert!(head.starts_with("⚠ conflicting host: Sunshine (running), Apollo"));
+    fn a_summary_carrying_conflicts_still_deserializes_and_is_ignored() {
+        let json = r#"{"version":"0.5.1","video_streaming":false,"audio_streaming":false,
+            "session":null,"paired_clients":1,"native_paired_clients":2,"pin_pending":false,
+            "pending_approvals":0,"kept_displays":0,"conflicts":["Sunshine (installed)"]}"#;
+        let s: Summary = serde_json::from_str(json).expect("unknown fields are ignored");
+        assert_eq!(
+            TrayStatus::Running(s).headline(),
+            "punktfunk host 0.5.1 — idle"
+        );
    }

    #[test]
@@ -478,6 +485,32 @@ mod tests {
            .contains("status unavailable"));
    }

+    /// A live session means streaming even if the host's flag says otherwise — a host from before
+    /// the `get_local_summary` fix only raised `video_streaming` for the GameStream plane, so a
+    /// native session showed as "idle" with its own mode printed next to it.
+    #[test]
+    fn a_live_session_reads_as_streaming_without_the_flag() {
+        let mut s = summary(true);
+        s.video_streaming = false; // pre-fix host, native session
+        let st = TrayStatus::Running(s);
+        assert!(st.is_streaming());
+        assert_eq!(
+            st.headline(),
+            "punktfunk host 0.5.1 — streaming 2560×1440@120"
+        );
+        // No session and no flag is still idle.
+        assert!(!TrayStatus::Running(summary(false)).is_streaming());
+    }
+
+    #[test]
+    fn kept_displays_are_reported_for_the_release_action() {
+        assert_eq!(TrayStatus::Running(summary(false)).kept_displays(), 0);
+        let mut s = summary(false);
+        s.kept_displays = 2;
+        assert_eq!(TrayStatus::Running(s).kept_displays(), 2);
+        assert_eq!(TrayStatus::Degraded.kept_displays(), 0);
+    }
+
    #[test]
    fn pairing_attention_flags() {
        let mut s = summary(false);
@@ -51,6 +51,7 @@ const IDM_RESTART: usize = 0x0104;
 const IDM_LOGS: usize = 0x0105;
 const IDM_EXIT: usize = 0x0106;
 const IDM_PAIRING: usize = 0x0107;
+const IDM_DISPLAYS: usize = 0x0108;

 /// Icon resource ordinals (embedded by build.rs).
 fn icon_ordinal(status: &TrayStatus) -> u16 {
@@ -73,8 +74,9 @@ struct App {
    taskbar_created: u32,
    /// `punktfunk-host.exe` next to this exe (the installer lays both in `{app}`).
    host_exe: Option<std::path::PathBuf>,
-    /// The console answered the poller's live loopback probe — the "Open web console" entry is
-    /// shown iff opening it would actually work (repo-run consoles included, stopped ones not).
+    /// The console answered the poller's live loopback probe. Drives the label of the (always
+    /// present) "Open web console" entry, and whether a left-click on the icon opens the console
+    /// or falls back to showing the menu.
    web_console: AtomicBool,
    web_port: u16,
 }
@@ -305,14 +307,24 @@ fn show_menu(hwnd: HWND) {
        };
        add(IDM_HEADER, &status.headline(), true);
        let _ = AppendMenuW(menu, MF_SEPARATOR, 0, PCWSTR::null());
+        // The console entry is ALWAYS here — it is the reason most people open this menu, and
+        // left-clicking the icon is not a discoverable substitute. When the loopback probe says
+        // the console isn't answering the label says so, rather than the entry vanishing.
        if app().web_console.load(Ordering::SeqCst) {
            add(IDM_OPEN_WEB, "Open web console", false);
-            let _ = SetMenuDefaultItem(menu, IDM_OPEN_WEB as u32, 0);
-            if status.pairing_attention() {
-                add(IDM_PAIRING, "Approve pairing request…", false);
-            }
-            let _ = AppendMenuW(menu, MF_SEPARATOR, 0, PCWSTR::null());
+        } else {
+            add(IDM_OPEN_WEB, "Open web console (not responding)", false);
        }
+        let _ = SetMenuDefaultItem(menu, IDM_OPEN_WEB as u32, 0);
+        if status.pairing_attention() {
+            add(IDM_PAIRING, "Approve pairing request…", false);
+        }
+        match status.kept_displays() {
+            0 => {}
+            1 => add(IDM_DISPLAYS, "Release kept display…", false),
+            n => add(IDM_DISPLAYS, &format!("Release {n} kept displays…"), false),
+        }
+        let _ = AppendMenuW(menu, MF_SEPARATOR, 0, PCWSTR::null());
        if can_control {
            if startable {
                add(IDM_START, "Start host", false);
@@ -385,8 +397,13 @@ fn elevate_service(hwnd: HWND, verb: &str) {
    }
 }

-fn open_web_console(hwnd: HWND) {
-    shell_open(hwnd, &format!("https://localhost:{}", app().web_port));
+/// Open the web console at `path` ("" = dashboard). Deep links land the operator on the page the
+/// menu entry promised — the pairing queue, the virtual displays — instead of the dashboard.
+fn open_web_console(hwnd: HWND, path: &str) {
+    shell_open(
+        hwnd,
+        &format!("https://localhost:{}/{path}", app().web_port),
+    );
 }

 fn open_logs(hwnd: HWND) {
@@ -416,7 +433,7 @@ extern "system" fn wndproc(hwnd: HWND, msg: u32, wparam: WPARAM, lparam: LPARAM)
                WM_CONTEXTMENU => show_menu(hwnd),
                x if x == NIN_SELECT || x == NIN_KEYSELECT => {
                    if app.web_console.load(Ordering::SeqCst) {
-                        open_web_console(hwnd);
+                        open_web_console(hwnd, "");
                    } else {
                        show_menu(hwnd);
                    }
@@ -427,8 +444,9 @@ extern "system" fn wndproc(hwnd: HWND, msg: u32, wparam: WPARAM, lparam: LPARAM)
        }
        WM_COMMAND => {
            match (wparam.0) & 0xffff {
-                IDM_OPEN_WEB => open_web_console(hwnd),
-                IDM_PAIRING => open_web_console(hwnd),
+                IDM_OPEN_WEB => open_web_console(hwnd, ""),
+                IDM_PAIRING => open_web_console(hwnd, "pairing"),
+                IDM_DISPLAYS => open_web_console(hwnd, "displays"),
                IDM_START => elevate_service(hwnd, "start"),
                IDM_STOP => elevate_service(hwnd, "stop"),
                IDM_RESTART => elevate_service(hwnd, "restart"),
@@ -0,0 +1,88 @@
+NT-handle import: consume on SUCCESS only — PUNKTFUNK LOCAL PATCH.
+
+Not upstream. Granite's DeviceAllocator::internal_allocate closed the imported
+by-reference NT handle (OPAQUE_WIN32 / D3D11_TEXTURE / D3D12_RESOURCE — the
+predicate excludes fd and KMT types) UNCONDITIONALLY after the FIRST
+vkAllocateMemory — on success AND on failure — while every failure before the
+allocator (c-shim validation, vkCreateImage, find_memory_type) left the handle
+open. Both classes surface to pyrowave_image_create's caller as the same error
+code, so the caller could not know whether to close: punktfunk's import_plane
+closed on every failure and double-closed whenever the failure was at or after
+the allocate — and a recycled handle value makes that close an unrelated live
+handle's. The unconditional consume also broke the allocator's own
+block-recycling retry loop, which re-ran vkAllocateMemory with import_info
+still pointing at the just-closed handle.
+
+Fixed by moving the consume to the API commit point: pyrowave_image_create
+closes the by-reference handle exactly on its success return. This pins
+pyrowave.h's documented contract ("take ownership and close the HANDLE on
+import") to mean on SUCCESS, matching pyrowave_sync_object_create (Granite's
+semaphore import already closes only after a successful vkImportSemaphore*).
+NT-handle memory imports never transfer ownership at the Vulkan level (the
+implementation holds its own reference), so the post-create close is legal.
+
+Behavior note: Granite code that imports by-reference memory WITHOUT going
+through pyrowave_image_create — in this vendored subset only wsi_dxgi.cpp's
+interop swapchain, which the pyrowave C API never reaches — now keeps the
+caller's handle open on allocate-stage failure instead of consuming it. That
+matches the semaphore contract; those callers own their close-on-failure.
+
+Correctness re-validated by pyrowave_win_smoke on real hardware after the
+change.
+
+diff --git a/crates/pyrowave-sys/vendor/pyrowave/Granite/vulkan/memory_allocator.cpp b/crates/pyrowave-sys/vendor/pyrowave/Granite/vulkan/memory_allocator.cpp
+index f33bdac3..fd035b66 100644
+--- a/crates/pyrowave-sys/vendor/pyrowave/Granite/vulkan/memory_allocator.cpp
+++ b/crates/pyrowave-sys/vendor/pyrowave/Granite/vulkan/memory_allocator.cpp
+@@ -732,16 +732,15 @@ bool DeviceAllocator::internal_allocate(
+ 		res = table->vkAllocateMemory(device->get_device(), &info, nullptr, &device_memory);
+ 	}
+ 
+-	// If we're importing, make sure we consume the native handle.
+-	if (external && bool(*external) &&
+-	    ExternalHandle::memory_handle_type_imports_by_reference(external->memory_handle_type))
+-	{
+-#ifdef _WIN32
+-		::CloseHandle(external->handle);
+-#else
+-		::close(external->handle);
+-#endif
+-	}
+	// PUNKTFUNK (patch 0006): the consume of by-reference native handles used to happen HERE,
+	// unconditionally — success AND failure of the first vkAllocateMemory. That made the caller's
+	// failure contract ambiguous (an allocate-stage failure had already closed the handle while a
+	// create/find_memory_type failure had not), and the block-recycling retry loop below re-ran
+	// vkAllocateMemory with import_info still pointing at the just-closed handle. The consume now
+	// lives at the API commit point (pyrowave_c.cpp: pyrowave_image_create's success return), so
+	// the allocator never closes a caller's handle and retries import a still-open handle.
+	// (fd-type imports were never affected: memory_handle_type_imports_by_reference excludes
+	// OPAQUE_FD/DMA_BUF, whose ownership vkAllocateMemory itself transfers on success.)
+ 
+ 	if (res == VK_SUCCESS)
+ 	{
+diff --git a/crates/pyrowave-sys/vendor/pyrowave/pyrowave_c.cpp b/crates/pyrowave-sys/vendor/pyrowave/pyrowave_c.cpp
+index eb917e64..985cd0a9 100644
+--- a/crates/pyrowave-sys/vendor/pyrowave/pyrowave_c.cpp
+++ b/crates/pyrowave-sys/vendor/pyrowave/pyrowave_c.cpp
+@@ -592,6 +592,21 @@ pyrowave_result pyrowave_image_create(const pyrowave_image_create_info *info, py
+ 	if (!img)
+ 		return PYROWAVE_ERROR_FAILED_EXTERNAL_HANDLE;
+ 
+	// PUNKTFUNK (patch 0006): consume the imported by-reference NT handle exactly here, at the
+	// API's success boundary — the allocator no longer closes it (see memory_allocator.cpp).
+	// This pins pyrowave.h's documented contract ("take ownership and close the HANDLE on
+	// import") to mean ON SUCCESS, matching pyrowave_sync_object_create's semantics: on ANY
+	// failure return the caller still owns the handle and can close it unconditionally.
+	// By-reference NT-handle imports never transfer ownership at the Vulkan level (the
+	// implementation keeps its own reference), so a post-create close here is always legal.
+#ifdef _WIN32
+	if (info->external_handle &&
+	    ExternalHandle::memory_handle_type_imports_by_reference(info->handle_type))
+	{
+		::CloseHandle(reinterpret_cast<HANDLE>(info->external_handle));
+	}
+#endif
+
+ 	auto *image = new pyrowave_image_opaque();
+ 	image->device = &device;
+ 	image->img = std::move(img);
@@ -732,16 +732,15 @@ bool DeviceAllocator::internal_allocate(
 		res = table->vkAllocateMemory(device->get_device(), &info, nullptr, &device_memory);
 	}

-	// If we're importing, make sure we consume the native handle.
-	if (external && bool(*external) &&
-	    ExternalHandle::memory_handle_type_imports_by_reference(external->memory_handle_type))
-	{
-#ifdef _WIN32
-		::CloseHandle(external->handle);
-#else
-		::close(external->handle);
-#endif
-	}
+	// PUNKTFUNK (patch 0006): the consume of by-reference native handles used to happen HERE,
+	// unconditionally — success AND failure of the first vkAllocateMemory. That made the caller's
+	// failure contract ambiguous (an allocate-stage failure had already closed the handle while a
+	// create/find_memory_type failure had not), and the block-recycling retry loop below re-ran
+	// vkAllocateMemory with import_info still pointing at the just-closed handle. The consume now
+	// lives at the API commit point (pyrowave_c.cpp: pyrowave_image_create's success return), so
+	// the allocator never closes a caller's handle and retries import a still-open handle.
+	// (fd-type imports were never affected: memory_handle_type_imports_by_reference excludes
+	// OPAQUE_FD/DMA_BUF, whose ownership vkAllocateMemory itself transfers on success.)

 	if (res == VK_SUCCESS)
 	{
@@ -592,6 +592,21 @@ pyrowave_result pyrowave_image_create(const pyrowave_image_create_info *info, py
 	if (!img)
 		return PYROWAVE_ERROR_FAILED_EXTERNAL_HANDLE;

+	// PUNKTFUNK (patch 0006): consume the imported by-reference NT handle exactly here, at the
+	// API's success boundary — the allocator no longer closes it (see memory_allocator.cpp).
+	// This pins pyrowave.h's documented contract ("take ownership and close the HANDLE on
+	// import") to mean ON SUCCESS, matching pyrowave_sync_object_create's semantics: on ANY
+	// failure return the caller still owns the handle and can close it unconditionally.
+	// By-reference NT-handle imports never transfer ownership at the Vulkan level (the
+	// implementation keeps its own reference), so a post-create close here is always legal.
+#ifdef _WIN32
+	if (info->external_handle &&
+	    ExternalHandle::memory_handle_type_imports_by_reference(info->handle_type))
+	{
+		::CloseHandle(reinterpret_cast<HANDLE>(info->external_handle));
+	}
+#endif
+
 	auto *image = new pyrowave_image_opaque();
 	image->device = &device;
 	image->img = std::move(img);
@@ -40,8 +40,16 @@ depends on the display manager driving the autologin:

 - **SDDM** (Bazzite, SteamOS): handled automatically — no setup.
 - **plasmalogin** (Nobara) and other display managers: the host must stop the display manager
-  itself for the length of the stream and restart it afterwards, which needs privilege. Allow it
-  with a polkit rule (adjust the unit and user names to your box):
+  itself for the length of the stream and restart it afterwards, which needs privilege. The
+  packages ship that privilege: a root helper (`/usr/libexec/punktfunk/pf-dm-helper`) behind its
+  own polkit action (`io.unom.punktfunk.dm-helper`), invoked automatically when the plain
+  `systemctl` verbs are denied — no setup. The helper only stops/restores the unit the
+  `display-manager.service` symlink points at, the same class of local-seat operation these
+  distros already authorize for their own session switcher (Nobara's `os-session-select`).
+
+  Installed from a tarball, or prefer not to ship the `allow_any` action? Remove the `.policy`
+  file and use a polkit rule scoped to your user instead (adjust the unit and user names to your
+  box) — the host tries the plain verbs first, so the rule takes precedence:

  ```js
  // /etc/polkit-1/rules.d/49-punktfunk-dm.rules
@@ -54,15 +62,16 @@ depends on the display manager driving the autologin:
  });
  ```

-  Without the rule the host degrades safely: it **attaches** to the live Gaming Mode session
-  instead (Game Mode stays on the box's display, mirrored to the client) rather than risk the
-  display manager. If the display-manager restart ever loses its privilege mid-restore,
-  `PUNKTFUNK_RECOVER_SESSION_CMD` (see [Configuration](/docs/configuration)) is fired as the
-  fallback.
+  With no privilege path at all the host degrades safely: it **attaches** to the live Gaming Mode
+  session instead (Game Mode stays on the box's display at the box's own resolution, mirrored to
+  the client — if your monitor stays on and the stream runs at the desktop's resolution, this is
+  what happened; check the host log for "managed takeover unavailable"). If the display-manager
+  restart ever loses its privilege mid-restore, `PUNKTFUNK_RECOVER_SESSION_CMD` (see
+  [Configuration](/docs/configuration)) is fired as the fallback.

-  With the rule in place the **in-stream session switch round-trips** in managed mode: Steam's
-  "Switch to Desktop" inside the streamed Game Mode returns the box to its desktop session and the
-  stream follows it there; the desktop's "Return to Gaming Mode" switches it forward again.
+  With the takeover authorized the **in-stream session switch round-trips** in managed mode:
+  Steam's "Switch to Desktop" inside the streamed Game Mode returns the box to its desktop session
+  and the stream follows it there; the desktop's "Return to Gaming Mode" switches it forward again.

 ## Session following

@@ -66,7 +66,7 @@ the Windows specifics follow.

 The installer also sets up the **web management console** (status, paired devices, the PIN pairing
 flow): it bundles the console plus its own runtime and runs it as the **`PunktfunkWeb`** task on
-**`https://<this-PC>:47992`**, starting at boot.
+**`https://<this-PC>:47992`**, starting at boot and at sign-in.

 #### Console login password

@@ -19,13 +19,24 @@ pkgbase=punktfunk
 # host+client with no JS tooling — mirroring the RPM spec's `%bcond_with web` / `%bcond_with
 # scripting` (off by default). Set PF_WITH_WEB=1 / PF_WITH_SCRIPTING=1 to also build them (each
 # appends its pkgname + bun to makedepends below).
-pkgname=('punktfunk-host' 'punktfunk-client')
+# aarch64 (Arch Linux ARM) is CLIENT-ONLY — see the `arch` note below. Dropping the host from
+# `pkgname` (rather than giving package_punktfunk-host a per-package arch) is what keeps makepkg
+# from entering the host's build/package path at all, and mirrors how PF_WITH_WEB extends this
+# array further down.
+if [ "${CARCH:-x86_64}" = 'aarch64' ]; then
+  pkgname=('punktfunk-client')
+else
+  pkgname=('punktfunk-host' 'punktfunk-client')
+fi
 # CI (.gitea/workflows/arch.yml) drives the version: stable tags -> X.Y.Z-1, main pushes ->
 # X.Y.Z-0.<run#> in the separate punktfunk-canary repo (mirrors the RPM's 0.ciN release; pkgrel
 # allows only digits+dots, so the run number carries the monotonic ordering).
 pkgver="${PF_PKGVER:-0.7.0}"
 pkgrel="${PF_PKGREL:-1}"
-arch=('x86_64')
+# The CLIENT builds and runs on aarch64 (Arch Linux ARM); the HOST does not — its encode stack is
+# NVENC/QSV/AMF, all x86. On aarch64 the split below drops punktfunk-host from `pkgname`, so
+# makepkg builds only the client and never tries to compile the host crate.
+arch=('x86_64' 'aarch64')
 url="https://git.unom.io/unom/punktfunk"
 license=('MIT OR Apache-2.0')
 # !lto: makepkg's `lto` option injects -flto=auto into CFLAGS; aws-lc-sys (rustls' crypto)
@@ -83,6 +94,10 @@ build() {
  # real RFI (clean P-frame recovery anchor via DPB reference slots; design/linux-vulkan-video-encode.md).
  # Pure Rust `ash` (no new lib, no link-time deps); default on for HEVC (PUNKTFUNK_VULKAN_ENCODE=0
  # opts back to libav VAAPI), and a failed open falls back to VAAPI so unsupported devices are safe.
+  if [ "$CARCH" = 'aarch64' ]; then
+    # Client-only: no host crate, so none of the encode features below apply.
+    cargo build --release --locked -p punktfunk-client-linux -p punktfunk-client-session
+  else
  cargo build --release --locked --features punktfunk-host/nvenc,punktfunk-host/vulkan-encode \
    -p punktfunk-host -p punktfunk-client-linux -p punktfunk-client-session
  # The status tray in its OWN cargo invocation — load-bearing, not tidiness. Cargo unifies features
@@ -91,6 +106,7 @@ build() {
  # runtime by design) then panics at startup ("there is no reactor running, must be called from the
  # context of a Tokio 1.x runtime"). Built alone, its zbus stays on async-io. (Same split the .deb does.)
  cargo build --release --locked -p punktfunk-tray
+  fi
  # Management web console (opt-in): the Nitro `bun`-preset .output bundle (Bun.serve TLS),
  # built AND run with bun.
  if [ "${PF_WITH_WEB:-0}" = 1 ]; then
@@ -129,6 +145,14 @@ package_punktfunk-host() {
  install -Dm0755 "$T/punktfunk-host"                 "$pkgdir/usr/bin/punktfunk-host"
  # /dev/uinput + /dev/uhid -> input group (virtual gamepads + DualSense UHID)
  install -Dm0644 "$R/scripts/60-punktfunk.rules"     "$pkgdir/usr/lib/udev/rules.d/60-punktfunk.rules"
+  # Managed gamescope takeover on DM-autologin boxes: root helper + polkit action so the host can
+  # stop/restore the display manager for the stream. Arch has no /usr/libexec — install under
+  # /usr/lib/punktfunk and rewrite the policy's exec.path annotation to match (the host probes both).
+  install -Dm0755 "$R/scripts/pf-dm-helper"           "$pkgdir/usr/lib/punktfunk/pf-dm-helper"
+  install -Dm0644 "$R/scripts/io.unom.punktfunk.dm-helper.policy" \
+    "$pkgdir/usr/share/polkit-1/actions/io.unom.punktfunk.dm-helper.policy"
+  sed -i 's#/usr/libexec/punktfunk/pf-dm-helper#/usr/lib/punktfunk/pf-dm-helper#' \
+    "$pkgdir/usr/share/polkit-1/actions/io.unom.punktfunk.dm-helper.policy"
  # vhci-hcd autoload — usbip transport for the virtual Steam Deck pad (Steam only adopts USB pads)
  install -Dm0644 "$R/scripts/punktfunk-modules.conf" "$pkgdir/usr/lib/modules-load.d/punktfunk.conf"
  # 32 MB UDP socket buffers (send-side headroom at high bitrate)
@@ -70,6 +70,25 @@ makepkg -si
 # …add the web console too (needs bun / bun-bin):
 PF_WITH_WEB=1 PF_SRCDIR="$(git rev-parse --show-toplevel)" makepkg -f --holdver
 ```
+
+### aarch64 (Arch Linux ARM) — the client
+
+The PKGBUILD declares `arch=('x86_64' 'aarch64')`. On aarch64 it builds the **client only** —
+`pkgname` drops `punktfunk-host`, so makepkg never enters the host's build or package path, and
+`build()` skips the host/tray cargo invocations and their NVENC/Vulkan-encode features. The host
+stays x86-only because its encode stack (NVENC/QSV/AMF) is.
+
+Nothing else changes — run the same command on an Arch Linux ARM box:
+
+```sh
+cd packaging/arch
+PF_SRCDIR="$(git rev-parse --show-toplevel)" makepkg -f --holdver
+# -> punktfunk-client-<ver>-<rel>-aarch64.pkg.tar.zst   (no punktfunk-host package)
+```
+
+There is no cross-compile path here: makepkg builds for `CARCH`, so this wants a real aarch64
+Arch machine (or an emulated Arch Linux ARM container, which is slow). Unlike the deb, it has
+**not** been verified end to end yet — there is no official arm64 Arch container to test in.
 Then the standard first-run (printed by the install scriptlet):
 ```sh
 sudo usermod -aG input "$USER"          # virtual gamepads; re-login after
@@ -171,3 +171,25 @@ docker run --rm -v "$PWD:/src" -w /src pf-noble \

 `BUNDLE_FFMPEG=1` needs `patchelf` and an FFmpeg install at `FFMPEG_PREFIX` (default `/opt/ffmpeg`,
 which the noble image provides).
+
+### The arm64 client `.deb`
+
+The **client** also ships for arm64 (`punktfunk-client_<version>_arm64.deb`, published to the same
+apt distribution — the registry keys pool entries by architecture, so an arm64 box needs no extra
+configuration). There is no arm64 **host** package: the Linux host encodes with NVENC/QSV/AMF, all
+x86.
+
+It is cross-compiled on an ordinary amd64 machine in `ci/rust-ci-arm64cross.Dockerfile` — the
+rust-ci toolchain plus an Ubuntu ports arm64 multiarch sysroot. No arm64 runner is involved:
+
+```sh
+docker build -f ci/rust-ci-arm64cross.Dockerfile -t pf-arm64cross .   # repo-root context
+docker run --rm -v "$PWD:/w" -w /w pf-arm64cross \
+  bash -lc 'VERSION=0.0.1 ARCH=arm64 TARGET=aarch64-unknown-linux-gnu \
+              bash packaging/debian/build-client-deb.sh'
+```
+
+`TARGET` moves the binaries to `target/<triple>/release`; `ARCH` sets the package's
+`Architecture:` field. Set both — one without the other builds an amd64 binary into a package
+labelled arm64, or vice versa. `dpkg-shlibdeps` reads the arm64 sonames straight out of the
+multiarch sysroot, so `Depends:` comes out right with no manual list.
@@ -7,24 +7,40 @@
 # the package names the target boxes ship. The client links no NVIDIA libs — no filter
 # needed.
 #
-# Usage: VERSION=0.0.1~ci42.gdeadbee [ARCH=amd64] bash packaging/debian/build-client-deb.sh
+# Usage: VERSION=0.0.1~ci42.gdeadbee [ARCH=amd64] [TARGET=<rust triple>] \
+#          bash packaging/debian/build-client-deb.sh
 # Output: dist/punktfunk-client_<version>_<arch>.deb
+#
+# TARGET cross-compiles (and moves the binaries under target/<triple>/release). Set it
+# together with ARCH — e.g. ARCH=arm64 TARGET=aarch64-unknown-linux-gnu inside the
+# ci/rust-ci-arm64cross.Dockerfile image, which carries the matching :arm64 sysroot.
 set -euo pipefail

 VERSION="${VERSION:?set VERSION (e.g. 0.0.1 or 0.0.1~ci42.gdeadbee)}"
 ARCH="${ARCH:-amd64}"
+TARGET="${TARGET:-}"
 PKG="punktfunk-client"
 CRATE="punktfunk-client-linux"
 ROOTDIR="$(cd "$(dirname "$0")/../.." && pwd)"
 cd "$ROOTDIR"

-BIN="target/release/$PKG"
+# Cargo drops a cross build under target/<triple>/release, a native one straight in
+# target/release.
+if [ -n "$TARGET" ]; then
+  OUTDIR="target/$TARGET/release"
+  CARGO_TARGET_ARGS=(--target "$TARGET")
+else
+  OUTDIR="target/release"
+  CARGO_TARGET_ARGS=()
+fi
+
+BIN="$OUTDIR/$PKG"
 # The Vulkan/Skia session streamer the shell execs for a connect — shipped alongside the shell
 # (the shell resolves it as its /usr/bin sibling), or desktop streaming breaks.
-SESSION_BIN="target/release/punktfunk-session"
+SESSION_BIN="$OUTDIR/punktfunk-session"
 if [ ! -x "$BIN" ] || [ ! -x "$SESSION_BIN" ]; then
-  echo "==> building $CRATE + punktfunk-client-session (release)"
-  cargo build --release --locked -p "$CRATE" -p punktfunk-client-session
+  echo "==> building $CRATE + punktfunk-client-session (release${TARGET:+ for $TARGET})"
+  cargo build --release --locked "${CARGO_TARGET_ARGS[@]}" -p "$CRATE" -p punktfunk-client-session
 fi

 STAGE="$(mktemp -d)"
@@ -52,6 +52,11 @@ SHAREDIR="$STAGE/usr/share/$PKG"
 # --- file layout (matches the RPM %install) ----------------------------------
 install -Dm0755 "$BIN"                              "$STAGE/usr/bin/$PKG"
 install -Dm0644 scripts/60-punktfunk.rules         "$STAGE/usr/lib/udev/rules.d/60-punktfunk.rules"
+# Managed gamescope takeover on DM-autologin boxes: root helper + polkit action so the host can
+# stop/restore the display manager for the stream (the helper derives the DM unit itself).
+install -Dm0755 scripts/pf-dm-helper               "$STAGE/usr/libexec/punktfunk/pf-dm-helper"
+install -Dm0644 scripts/io.unom.punktfunk.dm-helper.policy \
+                                                   "$STAGE/usr/share/polkit-1/actions/io.unom.punktfunk.dm-helper.policy"
 # vhci-hcd autoload — usbip transport for the virtual Steam Deck pad (Steam only adopts USB pads).
 install -Dm0644 scripts/punktfunk-modules.conf     "$STAGE/usr/lib/modules-load.d/punktfunk.conf"
 # UDP socket-buffer tuning (32 MB) — without it the kernel clamps the host's SO_SNDBUF to ~416 KB
@@ -125,6 +125,32 @@ curl -fsS --user "enricobuehler:$REGISTRY_TOKEN" \
 > home-worker-2 has no flatpak and no passwordless sudo to install it — so the Deck or the
 > privileged CI container are the only two viable build sites.

+### aarch64
+
+The manifest builds for aarch64 as well as x86_64. Two things are architecture-specific, and both
+are now expressed properly rather than hardcoded:
+
+* **`PKG_CONFIG_PATH`** contains the runtime's multiarch directory. flatpak-builder does *not*
+  shell-expand `env` values, so `${FLATPAK_ARCH}` would be taken literally — a `build-options.arch`
+  override supplies the aarch64 string instead, inheriting everything else.
+* **The prebuilt Skia archive** is per-target and pinned by sha256. There are now two `type: file`
+  sources discriminated by `only-arches`, both landing on the same `dest-filename`, so
+  `SKIA_BINARIES_URL` stays one literal path. Upstream publishes the aarch64 archive under the
+  same skia commit hash and the same resolved-feature key (`pdf-textlayout-vulkan`), so on a
+  skia-safe bump update both URLs and both hashes together.
+
+```sh
+ARCH=aarch64 bash packaging/flatpak/build-flatpak.sh
+# -> dist/punktfunk-client-<version>-aarch64.flatpak
+```
+
+`ARCH` defaults to this machine's, and the bundle name now carries the architecture so an x86_64
+and an aarch64 build can coexist in `dist/`. This is **not** a cross-compile: flatpak-builder runs
+the build in a sandbox for the target arch, so building aarch64 anywhere but an arm64 machine
+needs qemu binfmt and is very slow. Not yet verified end to end — the manifest is correct by
+construction and the Skia hash was checked against the published archive, but no aarch64 flatpak
+has been built.
+
 ## Manifest

 [`io.unom.Punktfunk.yml`](io.unom.Punktfunk.yml). Runtime `org.gnome.Platform//50`
@@ -16,6 +16,12 @@
 #                      iteration, non-reproducible). Default: offline (regenerates cargo-sources).
 #   BUILDER=...        override the flatpak-builder invocation (default: auto-detect host
 #                      flatpak-builder, else `flatpak run org.flatpak.Builder`).
+#   ARCH=...           target architecture (default: this machine's). `aarch64` builds the arm64
+#                      client — the manifest carries the per-arch PKG_CONFIG_PATH and the matching
+#                      prebuilt Skia archive. NOTE this is not a cross-compile: flatpak-builder
+#                      runs the build inside a sandbox for that arch, so building aarch64 on an
+#                      x86_64 box needs qemu binfmt registered and is very slow. Run it on an
+#                      arm64 machine.
 set -euo pipefail

 ROOTDIR="$(cd "$(dirname "$0")/../.." && pwd)"
@@ -25,7 +31,11 @@ APP_ID="io.unom.Punktfunk"
 MANIFEST="packaging/flatpak/io.unom.Punktfunk.yml"
 VERSION="${VERSION:-$(git describe --tags --always --dirty 2>/dev/null || echo 0.0.1-dev)}"
 VERSION="${VERSION#v}"
-BUNDLE="dist/punktfunk-client-${VERSION}.flatpak"
+# `flatpak --default-arch` reports flatpak's own name for this machine (x86_64 / aarch64).
+ARCH="${ARCH:-$(flatpak --default-arch 2>/dev/null || uname -m)}"
+# Arch-suffixed so an x86_64 and an aarch64 bundle can sit in dist/ together instead of the
+# second silently overwriting the first. (CI composes its own published filename.)
+BUNDLE="dist/punktfunk-client-${VERSION}-${ARCH}.flatpak"

 # --- pick a flatpak-builder (host binary, or the org.flatpak.Builder flatpak on the Deck) ---
 if [ -n "${BUILDER:-}" ]; then
@@ -72,18 +82,19 @@ else
 fi

 # --- build into a local ostree repo, then export a single-file bundle --------------------
-echo "==> flatpak-builder ($APP_ID, version $VERSION)"
+echo "==> flatpak-builder ($APP_ID, version $VERSION, arch $ARCH)"
 # --default-branch=stable matches CI / the hosted repo ref, so a locally-built install can also
 # track flatpak.unom.io. build-bundle must then be told the branch (else it defaults to `master`).
 "${FPB[@]}" --user --force-clean --disable-rofiles-fuse \
  --default-branch=stable \
+  --arch="$ARCH" \
  --install-deps-from=flathub \
  "${EXTRA_ARGS[@]}" \
  --repo="$ROOTDIR/.flatpak-repo" \
  "$ROOTDIR/.flatpak-build" "$MANIFEST"

 mkdir -p dist
-flatpak build-bundle "$ROOTDIR/.flatpak-repo" "$BUNDLE" "$APP_ID" stable
+flatpak build-bundle --arch="$ARCH" "$ROOTDIR/.flatpak-repo" "$BUNDLE" "$APP_ID" stable
 echo "built $BUNDLE"
 ls -lh "$BUNDLE"
 echo
@@ -121,6 +121,14 @@ build-options:
    LIBCLANG_PATH: /usr/lib/sdk/llvm20/lib
    # mold (shipped in rust-stable) speeds the ~450-crate link on the Deck APU.
    RUSTFLAGS: -C link-arg=-fuse-ld=mold
+  # The multiarch directory in PKG_CONFIG_PATH is per-architecture, and flatpak-builder does NOT
+  # shell-expand `env` values — so ${FLATPAK_ARCH} would be taken literally. An `arch` override is
+  # the supported way to vary it; everything else above is inherited. Only the runtime's own
+  # /usr/lib/<triple> changes, so aarch64 differs from x86_64 in exactly this one string.
+  arch:
+    aarch64:
+      env:
+        PKG_CONFIG_PATH: /app/lib/pkgconfig:/usr/lib/aarch64-linux-gnu/pkgconfig:/usr/lib/pkgconfig

 modules:
  # ---------------------------------------------------------------------------------------
@@ -240,7 +248,18 @@ modules:
      # crate's pinned skia commit + target + resolved features — when bumping skia-safe, build
      # the workspace once locally and take the new name from
      # target/release/build/skia-bindings-*/out/.cache/, then update url + sha256 together.
+      # One entry per architecture, selected by `only-arches`; both land on the SAME
+      # dest-filename, so SKIA_BINARIES_URL above stays a single literal path. The upstream
+      # release publishes the aarch64 archive under the same skia commit hash and the same
+      # resolved-feature key, so the two stay in lockstep on a skia-safe bump — take both new
+      # names and sha256s together.
      - type: file
+        only-arches: [x86_64]
        url: https://github.com/rust-skia/skia-binaries/releases/download/0.87.0/skia-binaries-e551f334ad5cbdf43abf-x86_64-unknown-linux-gnu-pdf-textlayout-vulkan.tar.gz
        sha256: b46e7061e6b9df792025acaf9b8b90180224c7cec63f4c3ce09af7ddddb8abfa
        dest-filename: skia-binaries.tar.gz
+      - type: file
+        only-arches: [aarch64]
+        url: https://github.com/rust-skia/skia-binaries/releases/download/0.87.0/skia-binaries-e551f334ad5cbdf43abf-aarch64-unknown-linux-gnu-pdf-textlayout-vulkan.tar.gz
+        sha256: 6994fb993064d7d4fff00bf7c79544ffc881bb5441b6768b31cc088540f4bef7
+        dest-filename: skia-binaries.tar.gz
@@ -123,3 +123,24 @@ docker run --rm -v "$PWD:/src" -w /src punktfunk-fedora-rpm \

 A plain `rpmbuild`/COPR build with no `pf_version`/`pf_release` defines produces `0.3.0-1` (the
 spec defaults).
+
+### aarch64 — the client RPM
+
+The **client** builds for aarch64; the **host** does not (its encode stack is NVENC/QSV/AMF, all
+x86). `PF_WITHOUT_HOST=1` drops the host binary, the tray, the headless-session data, the
+firewalld services and the main package's `%files`, leaving exactly one RPM: `punktfunk-client`.
+Omitting the main `%files` is what keeps rpm from emitting an empty `punktfunk` next to it.
+
+This is **not** a cross-compile — `%build` runs cargo for the host architecture, so run it on an
+arm64 machine (or an emulated arm64 container, which is very slow):
+
+```sh
+docker build --platform linux/arm64 -f ci/fedora-rpm.Dockerfile -t punktfunk-fedora-rpm-arm64 ci
+docker run --rm --platform linux/arm64 -v "$PWD:/src" -w /src punktfunk-fedora-rpm-arm64 \
+  bash -lc 'git config --global --add safe.directory /src && \
+            PF_VERSION=0.0.1 PF_WITHOUT_HOST=1 bash packaging/rpm/build-rpm.sh'
+# -> dist/punktfunk-client-0.0.1-1.fcNN.aarch64.rpm
+```
+
+`PF_WITHOUT_HOST=1` works on x86_64 too, if you only want the client RPM. The flag is orthogonal
+to the architecture; it is just that aarch64 has no other option.
@@ -19,6 +19,11 @@ WEB_OPT=()
 # requirement as web; default off so a bare `rpmbuild`/COPR still works.
 SCRIPTING_OPT=()
 [ "${PF_WITH_SCRIPTING:-0}" = "1" ] && SCRIPTING_OPT=(--with scripting)
+# PF_WITHOUT_HOST=1 drops the host binary, the tray and the main package, leaving only
+# punktfunk-client. This is what an aarch64 build uses: the client is portable, the host's encode
+# stack (NVENC/QSV/AMF) is x86. Harmless on x86_64 too, if you only want the client RPM.
+HOST_OPT=()
+[ "${PF_WITHOUT_HOST:-0}" = "1" ] && HOST_OPT=(--without host)
 ROOTDIR="$(cd "$(dirname "$0")/../.." && pwd)"
 cd "$ROOTDIR"

@@ -39,7 +44,7 @@ git archive --format=tar.gz --prefix="punktfunk-${PF_VERSION}/" \
 # undefined cuStreamCreateWithPriority/cuMemcpy2DAsync_v2/…; this regen supersedes it.) Defining
 # extra unused symbols is harmless; a missing one fails the link. Only when /usr/lib64 is writable
 # (CI image runs as root) — COPR/mock provides the real cuda-cudart-devel stub instead.
-if [ "$(id -u)" = 0 ] && [ -d /usr/lib64 ]; then
+if [ "${PF_WITHOUT_HOST:-0}" != "1" ] && [ "$(id -u)" = 0 ] && [ -d /usr/lib64 ]; then
  CU_SYMS="$(grep -rhoE '\bcu[A-Z][A-Za-z0-9_]*' crates/punktfunk-host/src/ | sort -u || true)"
  if [ -n "$CU_SYMS" ]; then
    STUB_C="$(mktemp --suffix=.c)"
@@ -55,7 +60,7 @@ fi
 # resolves them from RPMs. Our builder image provides the toolchain via rustup (so
 # rust-toolchain.toml's pinned channel works) and the -devel libs via dnf, neither of which
 # rpmbuild's RPM-level check sees — skip it; a genuinely missing dep fails the compile/link.
-rpmbuild -bb --nodeps "${WEB_OPT[@]}" "${SCRIPTING_OPT[@]}" \
+rpmbuild -bb --nodeps "${WEB_OPT[@]}" "${SCRIPTING_OPT[@]}" "${HOST_OPT[@]}" \
  --define "_topdir $TOP" \
  --define "pf_version ${PF_VERSION}" \
  --define "pf_release ${PF_RELEASE}" \
@@ -63,6 +68,12 @@ rpmbuild -bb --nodeps "${WEB_OPT[@]}" "${SCRIPTING_OPT[@]}" \

 mkdir -p dist
 find "$TOP/RPMS" -name '*.rpm' -exec cp -v {} dist/ \;
-echo "== Requires (must NOT contain libcuda) =="
-rpm -qp --requires dist/punktfunk-${PF_VERSION}-*.rpm 2>/dev/null | grep -iE 'cuda|nvidia' \
-  && echo "  !! NVIDIA/CUDA leak !!" || echo "  clean"
+# The libcuda leak check applies to the HOST package (the only thing that links the driver); a
+# client-only build has no such package to inspect.
+if [ "${PF_WITHOUT_HOST:-0}" = "1" ]; then
+  echo "== client-only build: no host RPM, libcuda check not applicable =="
+else
+  echo "== Requires (must NOT contain libcuda) =="
+  rpm -qp --requires dist/punktfunk-${PF_VERSION}-*.rpm 2>/dev/null | grep -iE 'cuda|nvidia' \
+    && echo "  !! NVIDIA/CUDA leak !!" || echo "  clean"
+fi
@@ -31,10 +31,10 @@ URL:            https://git.unom.io/unom/punktfunk
 # COPR SCM builds provide the checkout; for a tarball build, drop a git archive here:
 Source0:        %{name}-%{version}.tar.gz

-# punktfunk-host is Linux-only and links system FFmpeg/PipeWire/Opus. x86_64 only for now: encode
-# is NVENC (desktop NVIDIA) and no aarch64 build is produced/published by CI — claiming aarch64
-# here would advertise an arch we never ship. Re-add aarch64 once there's an arm64 build leg.
-ExclusiveArch:  x86_64
+# punktfunk-host is Linux-only and links system FFmpeg/PipeWire/Opus. The HOST is x86_64 only —
+# its encode stack is NVENC/QSV/AMF — but the CLIENT builds and runs fine on aarch64, so the spec
+# accepts both arches and `--without host` (below) selects the client-only build.
+ExclusiveArch:  x86_64 aarch64

 # The zerocopy FFI links the NVIDIA driver's libcuda.so.1; rpm's auto-dep generator would turn
 # that into a hard Requires on libcuda.so.1 (and we never want to pin the driver — NVENC/EGL come
@@ -55,6 +55,14 @@ ExclusiveArch:  x86_64
 # Debian punktfunk-scripting .deb.
 %bcond_with scripting

+# The HOST half of this spec (the punktfunk package itself + the tray). ON by default, so an
+# ordinary x86_64 build is unchanged. `--without host` drops the host binary, the tray, the
+# headless-session data, the firewalld services and the main %%files section entirely, leaving
+# only punktfunk-client — which is what an aarch64 build produces, since the host's encode stack
+# (NVENC/QSV/AMF) is x86 and the client's is not. Omitting the main %%files is what stops rpm
+# from emitting an empty `punktfunk` package alongside the client.
+%bcond_without host
+
 # --- Build toolchain ---------------------------------------------------------
 BuildRequires:  cargo
 BuildRequires:  rust
@@ -211,14 +219,21 @@ export PUNKTFUNK_BUILD_VERSION="%{version}-%{release}"
 # with real RFI (clean P-frame recovery anchor via DPB reference slots; design/linux-vulkan-video-encode.md).
 # Pure Rust `ash` (no new lib / no link-time dep); default on for HEVC (PUNKTFUNK_VULKAN_ENCODE=0 opts
 # back to libav VAAPI), and a failed open falls back to VAAPI so unsupported devices degrade gracefully.
+%if %{with host}
 cargo build --release --locked --features punktfunk-host/nvenc,punktfunk-host/vulkan-encode \
  -p punktfunk-host -p punktfunk-client-linux -p punktfunk-client-session
+%else
+# Client-only (aarch64): no host crate, so none of the encode features apply.
+cargo build --release --locked -p punktfunk-client-linux -p punktfunk-client-session
+%endif
 # The status tray in its OWN cargo invocation — load-bearing, not tidiness. Cargo unifies features
 # across everything in one build, so co-building the tray with the host pulls the host's
 # ashpd -> zbus/tokio onto the tray's shared zbus; the tray (ksni async-io + blocking, no tokio
 # runtime by design) then panics at startup ("there is no reactor running, must be called from the
 # context of a Tokio 1.x runtime"). Built alone, its zbus stays on async-io. (Same split the .deb does.)
+%if %{with host}
 cargo build --release --locked -p punktfunk-tray
+%endif

 %if %{with web}
 # Management web console: build the Nitro SSR bundle with bun (the `bun` preset + our Bun.serve
@@ -243,12 +258,19 @@ fi
 %endif

 %install
+%if %{with host}
 # Binary
 install -Dm0755 target/release/punktfunk-host %{buildroot}%{_bindir}/punktfunk-host

 # udev rule — /dev/uinput access for virtual gamepads (input group).
 install -Dm0644 scripts/60-punktfunk.rules %{buildroot}%{_udevrulesdir}/60-punktfunk.rules

+# Managed gamescope takeover on DM-autologin boxes (Nobara's plasmalogin): a root helper + polkit
+# action let the host stop/restore the display manager for the stream without a hand-installed
+# polkit rule. The helper derives the DM unit itself — callers can't name arbitrary units.
+install -Dm0755 scripts/pf-dm-helper %{buildroot}%{_libexecdir}/punktfunk/pf-dm-helper
+install -Dm0644 scripts/io.unom.punktfunk.dm-helper.policy %{buildroot}%{_datadir}/polkit-1/actions/io.unom.punktfunk.dm-helper.policy
+
 # vhci-hcd autoload — the usbip transport that makes the virtual Steam Deck controller a
 # real USB device (Steam Input only adopts those; the UHID fallback is invisible to Steam).
 install -Dm0644 scripts/punktfunk-modules.conf %{buildroot}%{_prefix}/lib/modules-load.d/punktfunk.conf
@@ -288,6 +310,7 @@ for sz in 22x22 48x48; do
    install -Dm0644 "$png" %{buildroot}%{_datadir}/icons/hicolor/$sz/apps/"$(basename "$png")"
  done
 done
+%endif

 # --- client subpackage ---
 install -Dm0755 target/release/punktfunk-client %{buildroot}%{_bindir}/punktfunk-client
@@ -305,6 +328,7 @@ install -Dm0644 scripts/70-punktfunk-client.rules \
 install -Dm0644 scripts/99-punktfunk-client-net.conf \
                %{buildroot}%{_prefix}/lib/sysctl.d/99-punktfunk-client-net.conf

+%if %{with host}
 # Headless session helpers + example config + OpenAPI doc (reference material).
 install -d %{buildroot}%{_datadir}/%{name}/headless
 install -Dm0755 scripts/headless/run-headless-kde.sh   %{buildroot}%{_datadir}/%{name}/headless/run-headless-kde.sh
@@ -338,6 +362,7 @@ install -Dm0644 packaging/linux/punktfunk-native.xml \
 # Web console opener (TCP 47992) — only meaningful with the web subpackage, opened deliberately.
 install -Dm0644 packaging/linux/punktfunk-web.xml \
                %{buildroot}%{_prefix}/lib/firewalld/services/punktfunk-web.xml
+%endif

 %if %{with web}
 # --- web console subpackage (punktfunk-web) ---
@@ -377,12 +402,16 @@ chmod 0755 %{buildroot}%{_bindir}/punktfunk-scripting
 install -Dm0644 scripts/punktfunk-scripting.service %{buildroot}%{_userunitdir}/punktfunk-scripting.service
 %endif

+%if %{with host}
 %files
 %license LICENSE-MIT LICENSE-APACHE THIRD-PARTY-NOTICES.txt
 %doc README.md packaging/README.md
 %{_bindir}/punktfunk-host
 %{_bindir}/punktfunk-tray
 %{_udevrulesdir}/60-punktfunk.rules
+%dir %{_libexecdir}/punktfunk
+%{_libexecdir}/punktfunk/pf-dm-helper
+%{_datadir}/polkit-1/actions/io.unom.punktfunk.dm-helper.policy
 %{_prefix}/lib/modules-load.d/punktfunk.conf
 %{_prefix}/lib/sysctl.d/99-punktfunk-net.conf
 %{_prefix}/lib/firewalld/services/punktfunk-gamestream.xml
@@ -398,6 +427,7 @@ install -Dm0644 scripts/punktfunk-scripting.service %{buildroot}%{_userunitdir}/
 %config(noreplace) /etc/gamescope-session-plus/sessions.d/steam
 %dir %{_datadir}/%{name}
 %{_datadir}/%{name}/*
+%endif

 %files client
 %license LICENSE-MIT LICENSE-APACHE THIRD-PARTY-NOTICES.txt
@@ -441,6 +471,7 @@ udevadm trigger --subsystem-match=hidraw 2>/dev/null || :
 # rpm-ostree it takes effect on the next boot into the layered deployment).
 sysctl -p %{_prefix}/lib/sysctl.d/99-punktfunk-client-net.conf >/dev/null 2>&1 || :

+%if %{with host}
 %post
 # Reload udev so /dev/uinput picks up the new rule without a reboot (best-effort).
 udevadm control --reload-rules 2>/dev/null || :
@@ -465,6 +496,7 @@ if command -v punktfunk-host >/dev/null 2>&1; then
        echo "$conflict"
    fi
 fi
+%endif

 %if %{with web}
 %post web
@@ -0,0 +1,25 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE policyconfig PUBLIC
+ "-//freedesktop//DTD PolicyKit Policy Configuration 1.0//EN"
+ "http://www.freedesktop.org/standards/PolicyKit/1.0/policyconfig.dtd">
+<policyconfig>
+  <vendor>Punktfunk</vendor>
+  <vendor_url>https://punktfunk.unom.io</vendor_url>
+
+  <!-- Stop/restore the box's display manager for a managed gamescope takeover. The helper derives
+       the unit from the display-manager.service symlink itself (no caller-chosen units), so this
+       grant is scoped to the box's own local-seat session lifecycle — the same class of operation
+       these distros already authorize for their session switcher (e.g. Nobara's
+       os-session-select, allow_any). allow_any because the host commonly runs sessionless (a
+       lingering user unit, no polkit agent), where interactive auth can never be answered. -->
+  <action id="io.unom.punktfunk.dm-helper">
+    <description>Stop or restore the display manager for a Punktfunk stream</description>
+    <message>Authentication is required to switch the display manager for a Punktfunk stream</message>
+    <defaults>
+      <allow_any>yes</allow_any>
+      <allow_inactive>yes</allow_inactive>
+      <allow_active>yes</allow_active>
+    </defaults>
+    <annotate key="org.freedesktop.policykit.exec.path">/usr/libexec/punktfunk/pf-dm-helper</annotate>
+  </action>
+</policyconfig>
@@ -0,0 +1,39 @@
+#!/bin/sh
+# Privileged display-manager verbs for the punktfunk managed gamescope takeover.
+#
+# On DM-autologin boxes whose display manager does not survive a masked session unit (Nobara's
+# plasmalogin, unknown DMs), taking the Gaming Mode session over means stopping the display
+# manager for the length of the stream and restarting it afterwards — root-only operations. The
+# host invokes this helper via pkexec under its own polkit action
+# (io.unom.punktfunk.dm-helper.policy, installed by the packages), the same mechanism these
+# distros use for their own session switching (Nobara's os-session-select).
+#
+# The unit is NEVER caller-controlled: it is derived here from the display-manager.service alias
+# symlink, so the polkit grant's blast radius is exactly "the box's own display manager" — a
+# local-seat operation, not arbitrary unit management.
+set -eu
+
+dm_unit() {
+  target=$(readlink /etc/systemd/system/display-manager.service) || {
+    echo "pf-dm-helper: no display-manager.service alias — no display manager to manage" >&2
+    exit 1
+  }
+  basename "$target"
+}
+
+case "${1-}" in
+  stop)
+    exec systemctl stop "$(dm_unit)"
+    ;;
+  restore)
+    dm=$(dm_unit)
+    # reset-failed first: a relogin loop may have tripped the unit's start limit, and a plain
+    # restart would be refused until the accounting is cleared.
+    systemctl reset-failed "$dm" 2>/dev/null || :
+    exec systemctl restart "$dm"
+    ;;
+  *)
+    echo "usage: pf-dm-helper stop|restore" >&2
+    exit 2
+    ;;
+esac
@@ -12,6 +12,11 @@ Description=punktfunk management web console
 # web-init generates the login password; the host writes the mgmt token. Order after both.
 After=punktfunk-web-init.service punktfunk-host.service
 Wants=punktfunk-web-init.service
+# Retry indefinitely while the host is still writing the mgmt token + identity cert. Without this,
+# systemd's default rate limit (5 starts / 10 s) plus RestartSec=2 gives up permanently after ~10 s
+# - so a console enabled before the host's first run stayed dead until someone restarted it by hand
+# (the same defect the Windows PunktfunkWeb task had).
+StartLimitIntervalSec=0

 [Service]
 Type=simple
@@ -16,17 +16,26 @@ set "CERTFILE=%PFDATA%\cert.pem"
 set "KEYFILE=%PFDATA%\key.pem"

 rem The host's `serve` writes the mgmt token + its identity cert/key on first run. Until they exist
-rem we have no credential and no TLS material, so fail and let the task's restart-on-failure retry
-rem (mirrors the Linux unit's Restart=on-failure waiting for the host to create them) rather than
-rem silently downgrading to plain HTTP.
-if not exist "%TOKENFILE%" (
-  echo [punktfunk-web] mgmt token not present yet at "%TOKENFILE%" - waiting for the host service.
-  exit /b 1
-)
-if not exist "%CERTFILE%" (
-  echo [punktfunk-web] host identity cert not present yet at "%CERTFILE%" - waiting for the host service.
+rem we have no credential and no TLS material, so WAIT rather than silently downgrading to plain HTTP.
+rem
+rem Wait in-process instead of exiting 1 and hoping the task's restart-on-failure retries: Task
+rem Scheduler does not reliably restart on a plain non-zero exit code, so a console that started
+rem before the host finished writing those files (the token lands at argument parse, the cert only
+rem after the RSA-2048 keygen) used to stay down until the next reboot. ~5 min at 2 s, then give up
+rem so a genuinely broken install still surfaces as a failed task rather than one that runs forever.
+rem `timeout` needs a console this task does not have, so `ping -n 3` is the 2-second sleep.
+set /a PFWAITS=0
+:pfwait
+if exist "%TOKENFILE%" if exist "%CERTFILE%" goto pfready
+if %PFWAITS% GEQ 150 (
+  echo [punktfunk-web] gave up waiting for "%TOKENFILE%" + "%CERTFILE%" - is the punktfunk host service running?
  exit /b 1
 )
+if %PFWAITS%==0 echo [punktfunk-web] waiting for the host service to write the mgmt token + identity cert...
+set /a PFWAITS+=1
+ping -n 3 127.0.0.1 >nul 2>&1
+goto pfwait
+:pfready

 rem Both files are single KEY=VALUE lines (LF), written 0600/ACL'd: PUNKTFUNK_MGMT_TOKEN=... and
 rem PUNKTFUNK_UI_PASSWORD=... . Split on the first '=' and import each into the environment.
@@ -54,7 +54,7 @@ export const LaunchSpec = Schema.Struct({ "kind": Schema.String.annotate({ "desc
 export type LayoutMode = "auto-row" | "manual"
 export const LayoutMode = Schema.Literals(["auto-row", "manual"]).annotate({ "description": "How group members are arranged in the desktop coordinate space. Stored at Stage 0; applied from\nthe multi-monitor stage." })
 export type LocalSummary = { readonly "audio_streaming": boolean, readonly "conflicts"?: ReadonlyArray<string>, readonly "kept_displays": number, readonly "native_paired_clients": number, readonly "paired_clients": number, readonly "pending_approvals": number, readonly "pin_pending": boolean, readonly "session"?: null | { readonly "fps": number, readonly "height": number, readonly "width": number }, readonly "version": string, readonly "video_streaming": boolean }
-export const LocalSummary = Schema.Struct({ "audio_streaming": Schema.Boolean.annotate({ "description": "True while the audio stream thread is running." }), "conflicts": Schema.optionalKey(Schema.Array(Schema.String).annotate({ "description": "Other Moonlight-compatible hosts (Sunshine/Apollo/…) detected on this machine at startup —\nrunning one alongside Punktfunk is unsupported. Compact labels (e.g. `Sunshine (running)`);\nthe tray/console surface them so the clash is visible before pairing silently fails." })), "kept_displays": Schema.Number.annotate({ "description": "Virtual displays being KEPT with no live session — lingering (keep-alive window) or pinned\n(`keep_alive: forever`). Non-zero means a display (and, exclusive, your physical monitors) is\nheld; the tray surfaces it + a one-click release. Active (in-use) displays are not counted.", "format": "int32" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)), "native_paired_clients": Schema.Number.annotate({ "description": "Number of paired native (punktfunk/1) devices.", "format": "int32" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)), "paired_clients": Schema.Number.annotate({ "description": "Number of pinned (paired) GameStream client certificates.", "format": "int32" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)), "pending_approvals": Schema.Number.annotate({ "description": "Native pairing knocks awaiting the operator's approval (count only).", "format": "int32" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)), "pin_pending": Schema.Boolean.annotate({ "description": "True while a GameStream pairing handshake is parked waiting for the user's PIN." }), "session": Schema.optionalKey(Schema.Union([Schema.Null, Schema.Struct({ "fps": Schema.Number.annotate({ "format": "int32" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)), "height": Schema.Number.annotate({ "format": "int32" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)), "width": Schema.Number.annotate({ "format": "int32" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)) }).annotate({ "description": "The active launch session (set by Moonlight's `/launch`, cleared on cancel/stop)." })], { mode: "oneOf" })), "version": Schema.String.annotate({ "description": "Host version (mirrors `/health`)." }), "video_streaming": Schema.Boolean.annotate({ "description": "True while the video stream thread is running." }) }).annotate({ "description": "Non-sensitive host status for the local tray icon: counts and booleans only — no PIN values,\nno fingerprints, no device names. Served unauthenticated to LOOPBACK peers only (see\n`require_auth`): the bearer-token file is SYSTEM/Administrators-DACL'd on Windows, so the\nper-user tray process cannot authenticate — this narrow read-only route is its status source." })
+export const LocalSummary = Schema.Struct({ "audio_streaming": Schema.Boolean.annotate({ "description": "True while audio is streaming on either plane (same rule as `video_streaming`)." }), "conflicts": Schema.optionalKey(Schema.Array(Schema.String).annotate({ "description": "Other Moonlight-compatible hosts (Sunshine/Apollo/…) detected on this machine at startup —\nrunning one alongside Punktfunk is unsupported. Compact labels (e.g. `Sunshine (running)`);\nthe tray/console surface them so the clash is visible before pairing silently fails." })), "kept_displays": Schema.Number.annotate({ "description": "Virtual displays being KEPT with no live session — lingering (keep-alive window) or pinned\n(`keep_alive: forever`). Non-zero means a display (and, exclusive, your physical monitors) is\nheld; the tray surfaces it + a one-click release. Active (in-use) displays are not counted.", "format": "int32" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)), "native_paired_clients": Schema.Number.annotate({ "description": "Number of paired native (punktfunk/1) devices.", "format": "int32" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)), "paired_clients": Schema.Number.annotate({ "description": "Number of pinned (paired) GameStream client certificates.", "format": "int32" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)), "pending_approvals": Schema.Number.annotate({ "description": "Native pairing knocks awaiting the operator's approval (count only).", "format": "int32" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)), "pin_pending": Schema.Boolean.annotate({ "description": "True while a GameStream pairing handshake is parked waiting for the user's PIN." }), "session": Schema.optionalKey(Schema.Union([Schema.Null, Schema.Struct({ "fps": Schema.Number.annotate({ "format": "int32" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)), "height": Schema.Number.annotate({ "format": "int32" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)), "width": Schema.Number.annotate({ "format": "int32" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)) }).annotate({ "description": "The active session: GameStream's launch (Moonlight `/launch`) when present, else the first\nlive native session. `null` when nothing is streaming." })], { mode: "oneOf" })), "version": Schema.String.annotate({ "description": "Host version (mirrors `/health`)." }), "video_streaming": Schema.Boolean.annotate({ "description": "True while video is streaming on EITHER plane: the GameStream media pipeline, or a live\nnative (punktfunk/1) session — the default plane, invisible in the GameStream flag alone." }) }).annotate({ "description": "Non-sensitive host status for the local tray icon: counts and booleans only — no PIN values,\nno fingerprints, no device names. Served unauthenticated to LOOPBACK peers only (see\n`require_auth`): the bearer-token file is SYSTEM/Administrators-DACL'd on Windows, so the\nper-user tray process cannot authenticate — this narrow read-only route is its status source." })
 export type LogEntry = { readonly "level": string, readonly "msg": string, readonly "seq": number, readonly "target": string, readonly "ts_ms": number }
 export const LogEntry = Schema.Struct({ "level": Schema.String.annotate({ "description": "`ERROR` | `WARN` | `INFO` | `DEBUG` | `TRACE`." }), "msg": Schema.String.annotate({ "description": "The formatted message, structured fields appended as `key=value`." }), "seq": Schema.Number.annotate({ "description": "Monotonic sequence number (1-based) — pass the last one back as the `after` cursor.", "format": "int64" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)), "target": Schema.String.annotate({ "description": "The emitting module path (tracing target)." }), "ts_ms": Schema.Number.annotate({ "description": "Unix timestamp in milliseconds.", "format": "int64" }).check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(0)) }).annotate({ "description": "One captured log event." })
 export type ModeConflict = "separate" | "steal" | "join" | "reject"
@@ -149,6 +149,14 @@ if (process.argv.includes("--list")) {
 	process.exit(0);
 }

+// Park the process instead of spinning it. `Effect.runFork` registers NOTHING with the event
+// loop — the keep-alive timer lives in `Runtime.makeRunMain`, which we deliberately don't use
+// (our shutdown is the bespoke two-signal one below) — and a bun process whose only pending work
+// is an unresolved promise BUSY-SPINS rather than blocking. With units running there is usually a
+// socket or timer to park on, but in the documented no-op state (no scripts, no plugins) there is
+// nothing at all: field report 2026-07-25 had it pinning a full core indefinitely, `strace`
+// showing a bare `clock_gettime` loop and nothing else. One idle handle is the whole fix.
+const keepAlive = setInterval(() => {}, 2 ** 31 - 1);
 const fiber = Effect.runFork(runner(options));
 let stopping = false;
 const shutdown = (signal: string) => {
@@ -161,3 +169,4 @@ process.on("SIGINT", () => shutdown("SIGINT"));
 process.on("SIGTERM", () => shutdown("SIGTERM"));

 await Effect.runPromise(Fiber.await(fiber));
+clearInterval(keepAlive); // every unit ended on its own — let the process exit
@@ -0,0 +1,56 @@
+// The shipped entry point's idle contract: `punktfunk-scripting` with nothing to run must PARK,
+// not spin. `Effect.runFork` registers nothing with the event loop (the keep-alive lives in
+// `Runtime.makeRunMain`, which runner-cli deliberately doesn't use), and a bun process whose only
+// pending work is an unresolved promise busy-loops instead of blocking — so in the runner's own
+// documented no-op state (no scripts, no plugins) it pinned a full core indefinitely until
+// runner-cli started holding one idle handle. Measured against the real CLI rather than mocked:
+// the regression is a property of the process, not of any function in it.
+import { afterAll, expect, test } from "bun:test";
+import * as fs from "node:fs";
+import * as path from "node:path";
+
+const SDK = path.join(import.meta.dir, "..");
+const ROOT = path.join(SDK, `.runner-cli-fixtures-${process.pid}`);
+const SCRIPTS = path.join(ROOT, "scripts");
+const PLUGINS = path.join(ROOT, "plugins");
+fs.mkdirSync(SCRIPTS, { recursive: true });
+fs.mkdirSync(PLUGINS, { recursive: true });
+afterAll(() => fs.rmSync(ROOT, { recursive: true, force: true }));
+
+// Long enough that bun's startup (~0.3 s of CPU, and it does not grow with the window) is a small
+// share of what we measure.
+const WINDOW_MS = 3_000;
+// A spinning runner burns a whole core — one window's worth of CPU in one window. A parked one
+// burns its startup and nothing after. Half the window sits far from both, so a contended CI
+// runner that only lets the spinner have half a core still fails.
+const MAX_CPU_MS = WINDOW_MS / 2;
+
+test(
+	"an idle runner parks instead of spinning",
+	async () => {
+		const proc = Bun.spawn(
+			[
+				process.execPath,
+				"src/runner-cli.ts",
+				"--scripts",
+				SCRIPTS,
+				"--plugins",
+				PLUGINS,
+			],
+			{ cwd: SDK, stdout: "pipe", stderr: "pipe" },
+		);
+		await Bun.sleep(WINDOW_MS);
+		proc.kill("SIGTERM");
+		await proc.exited;
+		const out = await new Response(proc.stdout).text();
+
+		// A process that died on startup also burns no CPU — assert it really ran the runner and
+		// was still alive to catch the signal, so the CPU bound below can't pass vacuously.
+		expect(out).toContain("[runner] nothing to run");
+		expect(out).toContain("[runner] SIGTERM");
+
+		const cpuMs = Number(proc.resourceUsage()?.cpuTime.total ?? 0) / 1000;
+		expect(cpuMs).toBeLessThan(MAX_CPU_MS);
+	},
+	WINDOW_MS + 15_000,
+);
@@ -18,16 +18,21 @@ set "CERTFILE=%PFDATA%\cert.pem"
 set "KEYFILE=%PFDATA%\key.pem"

 rem The host's `serve` writes the mgmt token + identity cert on first run. Until they exist the proxy
-rem has no credential and no TLS material, so fail and let restart-on-failure retry (mirrors the
-rem installed launcher / Linux unit) rather than silently serving plain HTTP.
-if not exist "%TOKENFILE%" (
-  echo [punktfunk-web] mgmt token not present yet at "%TOKENFILE%" - waiting for the host service.
-  exit /b 1
-)
-if not exist "%CERTFILE%" (
-  echo [punktfunk-web] host identity cert not present yet at "%CERTFILE%" - waiting for the host service.
+rem has no credential and no TLS material, so WAIT for them (mirrors the installed launcher) rather
+rem than silently serving plain HTTP - see scripts\windows\web-run.cmd for why waiting here beats
+rem exiting 1 and relying on the task's restart-on-failure. ~5 min at 2 s, then give up.
+set /a PFWAITS=0
+:pfwait
+if exist "%TOKENFILE%" if exist "%CERTFILE%" goto pfready
+if %PFWAITS% GEQ 150 (
+  echo [punktfunk-web] gave up waiting for "%TOKENFILE%" + "%CERTFILE%" - is the punktfunk host running?
  exit /b 1
 )
+if %PFWAITS%==0 echo [punktfunk-web] waiting for the host to write the mgmt token + identity cert...
+set /a PFWAITS+=1
+ping -n 3 127.0.0.1 >nul 2>&1
+goto pfwait
+:pfready

 rem Both files are single KEY=VALUE lines: PUNKTFUNK_MGMT_TOKEN=... and PUNKTFUNK_UI_PASSWORD=... .
 rem Split on the first '=' and import each into the environment.