feat(host/encode): de-escalate the latency escalation once cadence holds clean

Escalate-and-hold's missing half. The contention escalation (capture depth,
then the NVENC pipelined retrieve) was permanent: one sustained overrun — even
one CAUSED by the ABR overdrive's rebuild storms — cost the session its
depth-1 latency and its sub-frame streaming forever, and `encode_us` reported
queue depth instead of ASIC time for the rest of the session.

- The leaky bucket now keeps scoring after escalation. A sustained
  every-frame-on-cadence run (~5 s at 120 fps) winds back one stage in reverse
  order: pipelined retrieve first (its rebuild restores the IO-stream binding
  and sub-frame chunked streaming), then capture depth back to 1. Attempts are
  paced by an exponential backoff (1 → 5 → 25 min, capped) — a workload that
  truly needs the escalation converges to keeping it, but never a permanent
  latch.

- NVENC (Linux) implements `set_pipelined(false)`: a `want_sync` latch handled
  at the same drained safe point as the engage side (`maybe_disengage_async`
  mirrors `maybe_engage_async`); the lazy sync re-init re-arms everything and
  opens on an IDR. The stream loop polls until the switch lands, then re-runs
  the escalation warmup so the wind-back's own stall can't re-escalate it.
  `PUNKTFUNK_NVENC_ASYNC=1` (operator-pinned async) refuses the wind-back;
  the trait doc now specifies the two-way contract.

- While escalated, `cadence_degraded` stays latched (bitrate climbs refused)
  even with the bucket drained: the headroom is spent, and climbs resuming
  mid-escalation would saw against it and starve the clean run.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

.gitea/workflows/ci.yml

@@ -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
 

.gitea/workflows/windows-host.yml

@@ -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)

api/openapi.json

@@ -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."
           }
         }
       },

crates/pf-capture/src/linux/mod.rs

@@ -2652,6 +2652,18 @@ 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

crates/pf-encode/src/enc/codec.rs

@@ -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,
@@ -313,8 +352,15 @@ 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 +406,16 @@ 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
@@ -413,6 +469,17 @@ 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
@@ -454,6 +521,26 @@ 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(())
 }
 
@@ -477,6 +564,26 @@ 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] {

crates/pf-encode/src/enc/linux/mod.rs

@@ -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,
@@ -695,8 +724,10 @@ 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 +736,13 @@ 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 +758,7 @@ 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 +767,7 @@ 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 +776,9 @@ 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]);

crates/pf-encode/src/enc/linux/nvenc_cuda.rs

@@ -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, resolve_slices,
+    resolve_split_mode, resolve_subframe, store_ceiling, CeilingKey, LowLatencyConfig, NvStatusExt,
+    RFI_DPB,
 };
 use super::nvenc_status;
 use super::{AuChunk, ChromaFormat, Codec, EncodedFrame, Encoder, EncoderCaps};
@@ -595,6 +596,12 @@ 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 +709,7 @@ 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 +743,29 @@ 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() {
@@ -1000,6 +1031,23 @@ 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> {
@@ -1074,58 +1122,71 @@ 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 +1195,30 @@ 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(
+                        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 +1226,13 @@ 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 +1416,10 @@ 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 +1501,10 @@ 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);
@@ -1736,12 +1812,25 @@ 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();
@@ -2061,6 +2150,15 @@ 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 +2206,12 @@ 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 +2269,7 @@ mod tests {
             true,
             8,
             ChromaFormat::Yuv420,
+            false,
         )
         .expect("open NVENC CUDA session");
 
@@ -2241,6 +2346,7 @@ mod tests {
             true,
             8,
             ChromaFormat::Yuv444,
+            false,
         )
         .expect("open NVENC CUDA 4:4:4 session");
 
@@ -2286,6 +2392,7 @@ mod tests {
             true,
             8,
             ChromaFormat::Yuv420,
+            false,
         )
         .expect("open NVENC CUDA session");
 
@@ -2343,6 +2450,7 @@ mod tests {
             true,
             8,
             ChromaFormat::Yuv420,
+            false,
         ) else {
             eprintln!(
                 "skipping rfi_declines_impossible_ranges: NVENC unavailable (no NVIDIA driver)"
@@ -2369,6 +2477,7 @@ mod tests {
             true,
             8,
             ChromaFormat::Yuv420,
+            false,
         )
         .expect("open NVENC CUDA encoder")
     }
@@ -2403,6 +2512,7 @@ mod tests {
                 true,
                 8,
                 ChromaFormat::Yuv420,
+                false,
             )
             .expect("open");
             for f in 0..4u32 {
@@ -2540,6 +2650,7 @@ mod tests {
             true,
             8,
             ChromaFormat::Yuv420,
+            false,
         )
         .expect("open NVENC CUDA session");
         let frame = nv12_frame(W, H, 0);
@@ -2581,6 +2692,7 @@ mod tests {
             true,
             8,
             ChromaFormat::Yuv420,
+            false,
         )
         .expect("open NVENC CUDA session");
         // Steady sync frames first (stream-ordered mode).
@@ -2660,6 +2772,7 @@ mod tests {
             true,
             8,
             ChromaFormat::Yuv420,
+            false,
         )
         .expect("open NVENC CUDA session");
 
@@ -2765,6 +2878,7 @@ mod tests {
             true,
             8,
             ChromaFormat::Yuv420,
+            false,
         )
         .expect("open NVENC CUDA session");
 

crates/pf-encode/src/enc/linux/pyrowave.rs

@@ -227,12 +227,19 @@ impl PyroWaveEncoder {
         if !chroma.is_444() && (width % 2 != 0 || height % 2 != 0) {
             bail!("pyrowave 4:2:0 needs even dimensions (got {width}x{height})");
         }
-        if chroma.is_444() && !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. The 4:2:0 block
+        // count is ~half of 4:4:4's but still unbounded (8192×6144 4:2:0 = 73728 > u16::MAX), and
+        // the negotiator's 4:4:4 → 4:2:0 downgrade hands oversized modes to this open AS 4:2:0 —
+        // so a `chroma.is_444()`-gated check is skipped exactly when it is needed. Wrapping the
+        // index lets the resolve over-credit and `packetize` overshoot our bitstream buffer
+        // (its own bounds `assert` is compiled out by the Release vendored build).
+        // `validate_dimensions` rejects the impossible-at-any-chroma modes earlier; this is the
+        // 4:4:4-specific half plus defence in depth for the lab override.
+        if !crate::pyrowave_mode_fits_rdo(width, height, chroma.is_444()) {
             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 chroma.is_444() { "4:4:4" } else { "4:2:0" }
             );
         }
         // SAFETY: `open_inner` only issues Vulkan/pyrowave calls whose preconditions it

crates/pf-encode/src/enc/linux/vaapi.rs

@@ -39,11 +39,6 @@ use super::libav::{
 };
 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.

crates/pf-encode/src/enc/linux/vulkan_video.rs

@@ -221,6 +221,92 @@ impl NativeProfileStack {
 
 /// The Vulkan codec-operation bit for our codec selection (shared by open and the per-import
 /// profile rebuilds — the two must agree, profile identity is by value).
+/// The physical device + encode queue family a session runs on: the FIRST device exposing a
+/// `VIDEO_ENCODE` queue family that advertises `codec_op` (llvmpipe advertises none, so it drops
+/// out implicitly). Shared by [`VulkanVideoEncoder::open_inner`] and [`probe_encode_support`].
+///
+/// # Safety
+/// `instance` must be a live `ash::Instance` and `devices` handles enumerated from it.
+unsafe fn find_encode_device(
+    instance: &ash::Instance,
+    devices: &[vk::PhysicalDevice],
+    codec_op: vk::VideoCodecOperationFlagsKHR,
+) -> Option<(vk::PhysicalDevice, u32)> {
+    for &pd in devices {
+        let qf_len = instance.get_physical_device_queue_family_properties2_len(pd);
+        let mut video = vec![vk::QueueFamilyVideoPropertiesKHR::default(); qf_len];
+        let mut qf = vec![vk::QueueFamilyProperties2::default(); qf_len];
+        for i in 0..qf_len {
+            qf[i].p_next = &mut video[i] as *mut _ as *mut c_void;
+        }
+        instance.get_physical_device_queue_family_properties2(pd, &mut qf);
+        for i in 0..qf_len {
+            if qf[i]
+                .queue_family_properties
+                .queue_flags
+                .contains(vk::QueueFlags::VIDEO_ENCODE_KHR)
+                && video[i].video_codec_operations.contains(codec_op)
+            {
+                return Some((pd, i as u32));
+            }
+        }
+    }
+    None
+}
+
+/// Can this GPU + driver open a Vulkan Video **encode** session for `codec` at all?
+///
+/// This is the verdict the native-NV12 capture negotiation needs BEFORE it commits
+/// ([`crate::linux_native_nv12_ok`]): once the producer hands over two-plane NV12 there is no VAAPI
+/// fallback — libav would import it as packed RGB — so [`crate::open_video`] deliberately makes
+/// that open-failure fatal, and a Mesa built without `VK_KHR_video_encode_h265` therefore kills the
+/// session at its first frame instead of streaming on VAAPI.
+///
+/// Deliberately the FIRST check `open_inner` performs and hard-fails on, and nothing more: the same
+/// [`find_encode_device`] scan, run against the same codec op. That makes it provably no stricter
+/// than the open, so a failure here can only ever name a session that would have died anyway — it
+/// can never talk a working host out of the fast path. It is also cheap: one instance plus
+/// physical-device queries, no logical device, no video session, no VRAM. The later stages
+/// (`create_device`, `create_video_session`, the capability query) can still fail for reasons this
+/// does not model; those keep the session on the packed-RGB negotiation the ordinary way, because
+/// the capture format is only committed once this said yes.
+///
+/// Probed PER CODEC, not once: `codec_op_for` selects a different queue-family bit for AV1, and
+/// HEVC-encode-without-AV1-encode is the common VCN/ANV configuration.
+pub(crate) fn probe_encode_support(codec: Codec) -> Result<(), &'static str> {
+    if !matches!(codec, Codec::H265 | Codec::Av1) {
+        return Err("the Vulkan Video backend encodes HEVC + AV1 only");
+    }
+    let codec_op = codec_op_for(codec == Codec::Av1);
+    // SAFETY: creates one Vulkan instance and issues only physical-device queries against it, then
+    // destroys it on EVERY path below before returning — no handle derived from it escapes, and
+    // nothing outside this call observes it. `Entry::load` only dlopens the loader (a missing
+    // libvulkan returns `Err`), touching no process state the rest of the crate relies on.
+    // `find_encode_device` gets a live instance and handles enumerated from it, as it requires.
+    unsafe {
+        let Ok(entry) = ash::Entry::load() else {
+            return Err("no Vulkan loader");
+        };
+        let app = vk::ApplicationInfo::default().api_version(vk::API_VERSION_1_3);
+        let Ok(instance) = entry.create_instance(
+            &vk::InstanceCreateInfo::default().application_info(&app),
+            None,
+        ) else {
+            return Err("vkCreateInstance failed");
+        };
+        let found = match instance.enumerate_physical_devices() {
+            Ok(devices) => find_encode_device(&instance, &devices, codec_op).is_some(),
+            Err(_) => false,
+        };
+        instance.destroy_instance(None);
+        if found {
+            Ok(())
+        } else {
+            Err("no VK_KHR_video_encode queue for this codec on any device")
+        }
+    }
+}
+
 fn codec_op_for(av1: bool) -> vk::VideoCodecOperationFlagsKHR {
     if av1 {
         vk::VideoCodecOperationFlagsKHR::from_raw(
@@ -524,7 +610,9 @@ impl VulkanVideoEncoder {
     /// `open` with the RGB-direct request explicit instead of read from the env — the smoke
     /// tests use this (env mutation races parallel tests). `want_rgb` engages the RGB-direct
     /// source only if [`probe_rgb_direct`] also passes; otherwise the session opens on the CSC
-    /// path with the verdict logged.
+    /// path with the verdict logged. Test-only: the production entry point is [`Self::open`],
+    /// which resolves `want_rgb` from the env + the `cursor_blend` hint.
+    #[cfg(test)]
     pub(crate) fn open_opts(
         codec: Codec,
         width: u32,
@@ -602,34 +690,13 @@ impl VulkanVideoEncoder {
 
         let vq_inst = ash::khr::video_queue::Instance::new(&entry, &instance);
 
-        // pick the physical device + encode queue family (skip llvmpipe)
-        let (pd, encode_family) = {
-            let mut found = None;
-            for pd in instance.enumerate_physical_devices()? {
-                let qf_len = instance.get_physical_device_queue_family_properties2_len(pd);
-                let mut video = vec![vk::QueueFamilyVideoPropertiesKHR::default(); qf_len];
-                let mut qf = vec![vk::QueueFamilyProperties2::default(); qf_len];
-                for i in 0..qf_len {
-                    qf[i].p_next = &mut video[i] as *mut _ as *mut c_void;
-                }
-                instance.get_physical_device_queue_family_properties2(pd, &mut qf);
-                for i in 0..qf_len {
-                    if qf[i]
-                        .queue_family_properties
-                        .queue_flags
-                        .contains(vk::QueueFlags::VIDEO_ENCODE_KHR)
-                        && video[i].video_codec_operations.contains(codec_op)
-                    {
-                        found = Some((pd, i as u32));
-                        break;
-                    }
-                }
-                if found.is_some() {
-                    break;
-                }
-            }
-            found.context("no VK_KHR_video_encode queue for the requested codec on any device")?
-        };
+        // pick the physical device + encode queue family (skip llvmpipe). The scan itself lives in
+        // [`find_encode_device`] so the negotiation-time probe asks the SAME question this open
+        // asks — a probe that MIRRORS a dispatch goes stale the first time the dispatch grows a
+        // case, which is the failure `open_video`'s backend-label note already records.
+        let (pd, encode_family) =
+            find_encode_device(&instance, &instance.enumerate_physical_devices()?, codec_op)
+                .context("no VK_KHR_video_encode queue for the requested codec on any device")?;
         let mem_props = instance.get_physical_device_memory_properties(pd);
 
         // a compute queue family for the CSC (usually family 0) + its timestamp support (the

crates/pf-encode/src/enc/nvenc_core.rs

@@ -67,6 +67,155 @@ 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.
+
+    #[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`

crates/pf-encode/src/enc/nvenc_status.rs

@@ -72,9 +72,41 @@ 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"))
 }

crates/pf-encode/src/enc/pyrowave_wire.rs

@@ -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]

crates/pf-encode/src/enc/windows/amf.rs

@@ -1272,7 +1272,11 @@ 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
@@ -2760,6 +2764,10 @@ 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 +2786,8 @@ 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,

crates/pf-encode/src/enc/windows/ffmpeg_win.rs

@@ -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,
 }
@@ -151,8 +158,14 @@ 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 +284,11 @@ 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 +368,7 @@ 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 +490,10 @@ 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 +872,7 @@ 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 {

crates/pf-encode/src/enc/windows/nvenc.rs

@@ -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, resolve_slices,
+    resolve_split_mode, resolve_subframe, store_ceiling, CeilingKey, LowLatencyConfig, NvStatusExt,
+    RFI_DPB,
 };
 use super::nvenc_status;
 use super::{ChromaFormat, Codec, EncodedFrame, Encoder, EncoderCaps};
@@ -754,6 +755,27 @@ 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.
@@ -834,43 +856,14 @@ 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 +877,69 @@ 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 +948,15 @@ 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 +964,19 @@ 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)
-                            })
+                        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 +984,19 @@ 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;
+            // 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.)
@@ -1552,6 +1586,15 @@ 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 +1643,12 @@ 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.
     }

crates/pf-encode/src/enc/windows/pyrowave.rs

@@ -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);

crates/pf-encode/src/enc/windows/qsv.rs

@@ -774,7 +774,12 @@ 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)");
         }

crates/pf-encode/src/lib.rs

@@ -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)]
@@ -241,6 +243,12 @@ 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()
     }
@@ -853,18 +861,31 @@ 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 +894,42 @@ 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

crates/pf-vdisplay/src/vdisplay/linux/gamescope.rs

@@ -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() {

crates/pf-vdisplay/src/vdisplay/linux/gamescope/discovery.rs

@@ -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);
+    }
 }

crates/punktfunk-core/src/abr.rs

@@ -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;

crates/punktfunk-core/src/client/frame_channel.rs

@@ -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

crates/punktfunk-core/src/client/pump.rs

@@ -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,

crates/punktfunk-core/src/client/pump/control_task.rs

@@ -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");

crates/punktfunk-core/src/client/pump/data.rs

@@ -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"

crates/punktfunk-core/src/client/pump/datagram_task.rs

@@ -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);
                 }
             }

crates/punktfunk-host/src/mgmt/host.rs

@@ -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,
@@ -391,22 +393,23 @@ 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 {
+            native.first().map(|s| SessionInfo {
                 width: s.width,
                 height: s.height,
                 fps: s.fps,
@@ -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

crates/punktfunk-host/src/mgmt/tests.rs

@@ -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 = || {

crates/punktfunk-host/src/native.rs

@@ -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,

crates/punktfunk-host/src/native/control.rs

@@ -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,

crates/punktfunk-host/src/native/stream.rs

@@ -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;
+                        }
+                    }
                 }
             }
         }

crates/punktfunk-host/src/windows/install.rs

@@ -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));
     }
+    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]);
-    println!("web console set up + started (https://<host-ip>:47992)");
-    Ok(())
+        // 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<()> {

crates/punktfunk-tray/src/linux.rs

@@ -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();
     }
 }
@@ -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(),

crates/punktfunk-tray/src/status.rs

@@ -75,7 +75,7 @@ impl TrayStatus {
             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) {
+                let base = match (&s.session, self.is_streaming()) {
                     (Some(sess), true) => format!(
                         "punktfunk host {} — streaming {}×{}@{}",
                         s.version, sess.width, sess.height, sess.fps
@@ -113,8 +113,23 @@ impl TrayStatus {
         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)
+        matches!(self, TrayStatus::Running(s) if s.video_streaming || s.session.is_some())
+    }
+
+    /// 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 +171,11 @@ 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 +220,10 @@ 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 +240,13 @@ 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));
@@ -478,6 +505,32 @@ 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);

crates/punktfunk-tray/src/win.rs

@@ -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);
+        } 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);
         }
-            let _ = AppendMenuW(menu, MF_SEPARATOR, 0, PCWSTR::null());
+        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"),

docs-site/content/docs/gamescope.md

@@ -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
 

docs-site/content/docs/windows-host.md

@@ -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
 

packaging/arch/PKGBUILD

@@ -129,6 +129,14 @@ 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)

packaging/debian/build-deb.sh

@@ -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

packaging/rpm/punktfunk.spec

@@ -249,6 +249,12 @@ install -Dm0755 target/release/punktfunk-host %{buildroot}%{_bindir}/punktfunk-h
 # 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
@@ -383,6 +389,9 @@ install -Dm0644 scripts/punktfunk-scripting.service %{buildroot}%{_userunitdir}/
 %{_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

scripts/io.unom.punktfunk.dm-helper.policy

@@ -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>

scripts/pf-dm-helper

@@ -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

scripts/punktfunk-web.service

@@ -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

scripts/windows/web-run.cmd

@@ -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.

sdk/src/gen/punktfunk.ts

@@ -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"

web/web-run.cmd

@@ -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.