Opening a library source's settings did nothing, for every library plugin. Confirmed on
`.21` against the running console:
console origin :47992 /plugin-ui/lutris/__config -> 404
plugin origin :47993 /plugin-ui/lutris/__config -> 401
The drawer fetches a RELATIVE `/plugin-ui/<id>/__config`, so it resolves against the
console's own origin — where `middleware/auth.ts` answers 404 for `/plugin-ui/**`
unconditionally and by design. That refusal is the 2026-08-05 review's origin split
(H-3): plugin UIs moved to their own listener, and neither origin may serve the other's
paths. The drawer is the only consumer of `/plugin-ui` that is NOT an iframe — every
other caller builds an absolute URL from `pluginOriginFrom(uiConfig)` — so it was the
one thing the split broke, and nothing failed loudly enough to notice.
The fix is deliberately not to point the drawer at the plugin origin. That needs CORS
plus cross-site cookies, and it would put a plugin-controlled response inside a
credentialed cross-origin fetch — reopening exactly the hole the split closed. What
this drawer needs is DATA, not an embedded UI: `/api/plugin-config/<id>` reads the
plugin's `__config` server-side over loopback and returns the JSON same-origin, so no
plugin markup or script is ever served from the console origin and the per-boot secret
stays on the server, as with the `/plugin-ui` proxy.
`/api/**` is always session-gated (`isPublicPath`), so the new route inherits the gate
and answers 401 as JSON rather than redirecting to /login — which is what a `fetch`
needs and what the old path could never give it. It forwards only GET and PUT, reads
the body BEFORE the stale-credential retry (`readRawBody` drains the stream, so a
retried PUT would have saved `{}` over the operator's config), and passes the plugin's
own body through untouched so a 400's decode issue still reaches the operator.
Verified against the real built server: `/api/plugin-config/lutris` answers 401 — the
route resolves and is gated, and the BFF catch-all at `api/[...]` does not swallow it —
while `/plugin-ui/lutris/__config` still answers 404 on the console origin, i.e. the
split is intact. `/api/v1/status` still reaches the BFF. tsc clean, production build
clean, i18n 633 messages across en+de, biome clean on both touched files (the one
warning in SourceSettings.tsx pre-dates this change).
Four changes to the client interface, kept together because two of them touch the same rows
and the last is a bug the first would have made far more visible.
A thirteenth `ui_palette` entry, `oled`. The palette table is hand-mirrored in three languages
(`pf-console-ui`'s `library.rs`, `GamepadPalette.swift`, `GamepadPalette.kt`), so it goes into
all three at index 1, directly after the brand default — which keeps `PALETTES[0]` the unknown-id
fallback and keeps the dark-to-pale cycling order intact. What earns the name is arithmetic, not
a darker shade of violet: the ramp's first two stops are literally (0,0,0) and the ground is pure
black, so the shaded half of the field is pixels switched off rather than "very dark grey", and
the calm mix the form screens sit under lifts toward nothing at all. Mean cell luminance is 0.019
against Violet's 0.254. The bright corner keeps a faint indigo-to-violet ember so the backdrop is
still a field with somewhere to go, and that ember carries enough chroma at that luminance
(60 degrees of hue travel across 13 of the 16 cells) to satisfy the existing multi-tone assertion
without adding `oled` to the near-neutral exemption Graphite and Opal take. Each port gains an
`oled_is_actually_black` test that measures the claim — pure-black corner cells, a mean under half
the darkest other field's — rather than restating the table.
A new device key, `gamepad_ui_mode`. The gamepad-UI switch had been deciding two things at once:
whether to offer the controller-optimized interface at all, and that it appears only while a pad
is attached. A user asked for the second half to stop applying. `"connected"` (the default, and
exactly what the lone Bool meant) and `"always"` separate them, surfaced as a "Show it" row
directly under the switch on all five settings surfaces and built only while that switch is on —
a picker whose every option decides nothing is worse than no picker. `GamepadUIEnvironment.isActive`
takes the mode with NO default argument on purpose: a call site that forgot it would silently
strand everyone who chose Always back on "only with a controller", which is the one bug this
parameter exists to make impossible. An unrecognized value waits for a controller, so a mode a
newer client wrote can never trap an older one in a layout it has no way back out of. It stays a
device preference on both platforms, never part of a profile: which interface this device wears
has nothing to do with how a host streams to it.
The smoothness buffer is hidden under Lowest latency, not dimmed. Everywhere else already hid it
— the GTK and WinUI shells, the Apple touch and tvOS screens, the Android touch screen — because
under that intent it names a quantity that does not exist. Two surfaces disagreed: Apple's gamepad
settings screen left the row live and steppable, and the desktop console dimmed it, having no way
to drop a row from a fixed list. That list is now rebuilt each frame through a `row_applies`
filter. The concern about a vanishing row moving everything under the cursor does not apply here
and the new test says why: the row it drops sits directly BELOW the row that drops it, so the only
cursor that can be present when the list shrinks is the one on the intent row, which does not
move. Two latent hazards went with it — `apply_row` had been indexing the row list on the
assumption the cursor is always in range, and nothing re-clamped that cursor when another writer
changed the intent behind the screen's back.
Pale palettes were unreadable on tvOS, reported from the field. `GamepadInk` was never the
problem: it flips correctly for a pale field, it is not platform-gated, and every tvOS gamepad
entry point already published it. The cause is that this app sets `preferredColorScheme` nowhere
and declares no `UIUserInterfaceStyle`, so every SYSTEM-derived colour landing on those screens —
a `.secondary` placeholder, a `.bordered` button's chrome, a NavigationStack title, a material's
frost — resolved against the DEVICE appearance, which the palette cannot reach. On iPhone, iPad
and Mac a great many users sit in Light mode, so under a pale palette those colours came out dark
and the theme looked correct by accident; an Apple TV is Dark essentially always, so every one of
them rendered white on a light field. The mirror image was broken too and had simply never been
reported: a dark palette on a Light-mode iPhone was already drawing dark on dark. The scheme is
now published beside the ink, once, in `GamepadInkModifier`, because the two are halves of one
decision and publishing only the ink silently loses every colour the frameworks draw on the app's
behalf. Two structural amplifiers went with it: `ConsoleGlass` had been scoping the scheme to the
fill inside its `.background {}` on the tvOS and pre-26 branches while the 26 branch put it on the
content, so no console row's own content ever saw it on tvOS; and `LibraryView`'s navigation
chrome and its loading, error and empty states sit above `LibraryCoverflowView` and so were never
inked at all on tvOS and macOS, where that view is presented directly rather than through the
iOS-only `GamepadLibraryScreen` wrapper.
That last one exposed a second tvOS gap worth closing in the same breath: `ui_palette` had no row
in tvOS's ordinary Settings, and the gamepad settings screen that owns it everywhere else needs an
extended-profile controller to open on tvOS. An Apple TV driven by the Siri Remote alone could not
reach the palettes at all, which would now include the OLED one. `SettingsView.tvBody` carries a
Background row.
Verified: pf-console-ui builds, passes `clippy --all-targets -D warnings` and runs 74 tests clean
under linux/amd64 (a Mac `cargo check` of that crate is vacuous — every module is cfg'd to
linux/windows); `cargo fmt --check` clean for it and pf-client-core. Android `:app` runs 80 tests
with 0 failures, including four new `gamepadUiActive` cases and the palette parity table. The
Apple package builds for macOS AND tvOS and its 9 palette/gamepad-UI tests pass — the tvOS
typecheck is possible because the checked-in xcframework already carries a `tvos-arm64` slice. The
tvOS RENDERING fix is compile-verified only; an on-glass Apple TV check under a pale palette is
still owed, and is the one thing here that a build cannot answer.
`sdk-v0.1.3` failed at the publish step with `bun2nix: command not found`, exit 127.
Nothing was published, so 0.1.3 is still free.
`bun publish` runs the `prepare` lifecycle script, and sdk's `prepare` is
`bun2nix -o bun.nix` — regenerating the nix dependency file. That tool is a
devDependency of the repo, not something the `oven/bun:1` publish container has, and
the workflow's own install is `--ignore-scripts`, so nothing put it on PATH either.
This was latent, not new. `prepare` gained the bun2nix call on 2026-07-27 (1db8f763,
"move the bun packages to bun2nix"), while the last SDK publish was 0.1.2, bumped
2026-07-20. So the hook has been broken for every SDK release since it landed, and
0.1.3 is simply the first one to try. `@punktfunk/plugin-kit` has no `prepare` and was
never affected, which is why kit 0.3.2 published fine in that window and hid this.
The fix is NOT to copy `web/package.json`, which does the same job from `postinstall`.
That is right for web — it is never published — and would be worse here: a published
package's `postinstall` runs in every CONSUMER's install, so every plugin depending on
`@punktfunk/host` would try to run bun2nix and fail. `prepare` is the correct hook for
a published package (it does not run for consumers); it just must not assume a
repo-maintenance tool exists wherever a publish happens.
So the script skips when bun2nix is absent — and ONLY then. A present-but-failing
bun2nix still fails the script, because swallowing that would publish with a silently
stale bun.nix, which is the exact hand-maintained-hash problem 1db8f763 set out to end.
Both directions measured against the same `sh -e` bun and the Gitea runner use:
absent → exit 0, present-and-failing → exit 3.
`bun publish --dry-run` now completes and reports `+ @punktfunk/host@0.1.3`.
The library had one visibility control and it was all-or-nothing: turn a SOURCE off
and every one of its games goes. There was no way to drop a single title — a Proton
tool the filter missed, a demo, a game someone doesn't want on the TV — short of
hiding the whole launcher it came from.
**Where the setting lives.** Not on the entry. Only manual custom entries are stored;
a scanner's and a plugin's titles are rebuilt from scratch on every scan and every
reconcile, so a flag written onto one would be erased by the next sync — silently, and
minutes later, which is the worst possible shape for a setting. So `library-hidden.json`
holds the ids, mirroring how `library-scanners.json` holds disabled sources. The id is
stable by construction (D2: a claimed store's entries keep `<store>:<external_id>`
across reconciles), so a hide survives a re-scan, a plugin restart, and a store's
built-in→plugin migration.
**Where it takes effect.** In `all_games`, which is the one place every play surface
already funnels through — the grid on a client, native clients, the GameStream app
list, and launch resolution. Putting it there rather than at each call site is
deliberate: a per-surface filter is a rule someone has to remember, and forgetting one
is precisely the class of bug the `file://` art asymmetry in the previous commit was.
Hiding is curation, not access control — nothing is deleted, and un-hiding is instant.
**The console is the one surface that still sees them**, or a hidden title could never
be brought back. That exception is a TYPE, not a flag: `GET /library` answers
`Vec<GameEntry>` on every lane but the operator's and `Vec<OperatorGameEntry>` on
theirs, so a hidden entry cannot reach a paired streaming client by someone forgetting
a filter — there is no field there to leak. `hidden` is skipped when false, so the
response is byte-identical to today's for a library with nothing hidden.
`PUT /library/hidden/{id}` is operator-only — neither the plugin lane nor a paired cert,
unlike the scanner toggle. A plugin has no business deciding what its operator sees, and
a client must not be able to hide a game on the host it is streaming from. The id is not
validated against the current library on purpose: a title can be legitimately absent at
that moment (launcher closed, plugin mid-sync, drive unmounted), and refusing the
operator's choice in that window is worse than storing an id that matches nothing today.
On the card, the poster dims and a Hidden badge says why — a faded tile with no label
reads as a broken cover. Its controls stay at full contrast and, unlike an ordinary
card's, are not hover-revealed: the un-hide button is the only way out of the state, and
hiding it behind a hover would strand anyone on a touch screen.
Verified on .21 (Linux): 469 host tests pass (5 new), clippy clean under `-D warnings`,
`cargo fmt --all --check` clean. The routing test is the one that earns its keep — every
library id contains a colon and Heroic's contain two, so a router that split on it would
404 the console against ids the host itself produced. Console: tsc clean, production
build clean, i18n 633 messages across en+de, biome clean on the touched files.
The env-var reference had fallen behind the v0.25.0 CHANGELOG table. Added, with
the semantics taken from the code rather than the changelog one-liners:
- PUNKTFUNK_JUMBO / PUNKTFUNK_WIRE_MTU (Network & discovery), with a note
explaining the ack-gated mid-session grow, the start-at-1500 behavior, the
NIC/switch prerequisites, and the sub-1500 shrink direction of WIRE_MTU
- PUNKTFUNK_AUDIO_QUALITY / AUDIO_REDUNDANCY / AUDIO_OUTPUT_MODE — the legacy
HOST_AUDIO / KEEP_DEFAULT rows are folded into the OUTPUT_MODE row as the
aliases they now are (follow_default wins when both are set)
- PUNKTFUNK_NO_AUDIO_MINT (Windows minted-endpoint opt-out)
- PUNKTFUNK_PAD_AUDIO / PAD_AUDIO_SLOTS (Gamepads — DualSense speaker+haptics)
- PUNKTFUNK_NVENC_SPLIT_ARBITRATE (Advanced performance tuning)
- PUNKTFUNK_UI_PLUGIN_PORT / PUNKTFUNK_LIBRARY_ART_ROOTS (Auth, API & paths)
- PUNKTFUNK_VAAPI_DEVICE (client-side table)
Verified against the actual read sites (pf-host-config, wire_mtu.rs,
config.rs jumbo_wire_mtu, pad_audio.rs, minted.rs, art.rs, bun-https.mjs);
the page's remaining vars all still exist in code. MDX-compiles clean with GFM.
Three symptoms on .21, two defects. Lutris and Heroic appeared in the console sidebar
they explicitly opt out of; Lutris's settings were unreachable from the Library
screen; and Lutris and Steam logged `sync (startup) failed: HostRequestError`.
**The sidebar is a publish gap.** The console is correct — it keeps
`category: "library"` plugins out of the nav (`uiPlugins`, app-shell.tsx) — but the
host reports no category for them at all. `defineLibraryPlugin` sets it and
`sdk/src/ui.ts` forwards it; what SHIPS does not. `@punktfunk/host` was bumped to
0.1.2 on 2026-07-20 and `category` landed 2026-08-05 without a bump, so the registry's
0.1.2 is the pre-category build and every installed scanner registers without one.
Bumps the SDK to 0.1.3 — **inert until it is published**.
Because the field rides the untyped `pf.request` seam so an older host ignores it
rather than rejecting the registration, dropping it is silent by design. `serveUi` now
reads its own directory entry back and warns once when a requested category did not
land, the same way `defineLibraryPlugin` already warns when a store claim did not take.
That is what turns the next occurrence into a log line instead of a bug report.
**The missing settings and the failed sync are ONE defect: a write/read disagreement
about `file://`.** `local_art_bytes` decodes a `file://` value before testing
containment; `validate_art_paths` handed the raw value to `Path::new`, where
`file:///home/u/c.jpg` is a RELATIVE path whose first component is `file:`. It
canonicalized against the cwd, failed, and read as "outside every art root". So the
host refused every cover the kit's own `fileUrl` helper emits — the documented way for
a plugin to publish local art — while the read path would have served those same files.
That the two symptoms share a cause is not obvious and is why this is one commit: the
Library screen's settings control renders only for `origin: "plugin"`, and a source
becomes `plugin` only once it holds a store CLAIM, which is taken during a successful
reconcile. Lutris failed at entry 0 and Steam at entry 3, so neither ever claimed its
store, both stayed `origin: "builtin"`, and neither got a settings button. Heroic
reconciled (its art is http(s)) and has had its settings all along; rom-manager was
never affected because zero entries meant it never applied.
`art_path_is_servable` now decodes first, so both halves of the confinement judge the
same string. Confinement itself is unchanged: an out-of-root path is still refused in
`file://` clothing, which the test asserts alongside the accept case.
Diagnosing this took the HOST's journal, because both surfaces that should have
explained it lied. `HostRequestError` stringified to its bare tag, so the sync engine's
`${e.cause}` logged `HostRequestError` and discarded the method, the path and the
host's own message; it now renders all three, including an object-shaped cause that
used to print `[object Object]`. And the host logged "payload carries a field this lane
may not set" for BOTH refusals in `check_entry_fields`, so a 400 about an art path read
as an auth problem — it now logs the real reason and the entry title.
Verified on .21 (Linux): 463 host tests pass, clippy clean under `-D warnings`,
`cargo fmt --all --check` clean. The new art test fails without the fix and passes with
it. plugin-kit 71 and SDK 72 tests pass, both typecheck clean, biome clean.
A CachyOS / KDE Plasma 6.7.4 Wayland client with its 2560x1600@165 laptop panel at
150 % scaling negotiated 1706x1066 for "Native resolution" and streamed a visibly
blurry image. Two independent defects, and they stack — which is why forcing the mode
to 2560x1600 by hand did not fully fix it either.
1. `SDL_GetDesktopDisplayMode` reports a mode in SCREEN COORDINATES and hands the
pixels-per-point ratio back separately as `pixel_density`. We read `m.w`/`m.h` raw.
KDE advertises that panel as 1707x1067 points with a density of ~1.4997,
`render_scale::apply` even-floors both odd axes, and 1706x1066 goes on the wire —
exactly the mode in the reporter's handshake log. Multiplying by the density
recovers 2560x1600 to the pixel, because SDL derives it as the output's exact
pixels/points ratio. On X11 and Windows SDL never sets a density and `SDL_video.c`
normalizes the unset 0.0 to 1.0, so this is inert there: the bug needed a
compositor doing FRACTIONAL scaling.
2. The SDL window was created without `HIGH_PIXEL_DENSITY`, so the Wayland surface
stayed at buffer scale 1 — the Vulkan swapchain was built at 1707x1067 and KWin
upscaled it to the glass. Even a correct 2560x1600 stream was resampled down and
then back up. The same flaw silently shrank "Match window", which asks the host for
`size_in_pixels()`. The reporter's `SDL_VIDEO_WAYLAND_SCALE_TO_DISPLAY=1` workaround
is this same fix applied from outside SDL, which is why it helped.
The surrounding code was already written for pixels != points — the swapchain,
match-window and pointer mapping all read `size_in_pixels()` while window-size
persistence reads logical `size()` — so the flag only makes those two stop being the
same number. `display_scale()` starts reporting 1.5 into a swapchain that is 1.5x
larger, leaving the OSD the size it already was.
Also closes a smaller hole on the way past: only an `Err` from SDL reached the
1920x1080 fallback, so a display that reported a 0x0 mode sent a 0x0 request.
Verified on home-worker-5 (CachyOS — the reporter's distro, real SDL 3.4.14):
`cargo clippy --all-targets -p pf-presenter -- -D warnings` clean and 18/18
pf-presenter tests pass, three of them new and pinned to the field-reported numbers.
The 0.25.0 MacBook field report — audio jitter 'at certain points' — is the
jitter policy learning exclusively from audible failures, on both of its
sides. Growth needed THREE audible underruns before deepening the ring; the
A/V sync loop re-tested a shallower ring every five quiet seconds and paid an
audible starvation event every time it was wrong, forever; and a grown target
was never re-banked — growth raises a threshold, only a re-prime deepens the
ring — so a bunching link rode the knife edge, clicking once per bunching
period with the 'grown' target sitting inert. A ten-minute simulation of the
Wi-Fi power-save pattern (25 ms gaps / 300 ms, −50 ppm skew) measured ~2000
audible events under the shipped policy.
Three mechanisms, in JitterPolicy (Linux/Windows/Android) and mirrored in the
Swift AudioRing:
- NEAR-MISS: a read served with less than one protocol frame left over is the
same evidence as an underrun, heard by no one. It grows the target one step
per window, BEFORE the click — waiting for the third audible underrun means
the user heard two.
- SHRINK PROBES: every shrink is armed for five seconds; answered by an
underrun or near-miss it is undone on the spot, and a failed sync-driven
shrink is not retried for a doubling backoff (60 s → 8 min). A probe that
survives resets the backoff. Continuity outranks sync, now with a memory.
- HOLLOW RE-PRIME: an underrun while the depth AVERAGE runs more than a step
below the target re-primes immediately, spending the click it already cost
on the whole refill instead of limping. The average, not the instant, is
what separates a hollow ring from one late packet, and it is seeded on
prime so a fresh ring is never spuriously hollow.
Same simulation after: 9 audible events, tail clean but for the clock-skew
re-anchor (a genuinely slow host must re-bank every few minutes; only rate
adaptation would remove that, and no client has it). Neutralising the three
constants reproduces the ~2000 — the convergence tests fail against the old
behaviour.
Verified: 203 punktfunk-core tests, 254 Swift tests (5 skipped), clippy -D
warnings on punktfunk-core --all-features, cargo fmt --all --check.
The v0.25.0 rebuild published perfectly — registry has punktfunk-host 0.25.0-2 with
libavcodec.so=63-64, and it resolves on a real ffmpeg-9 box — then failed its last
step with
prune_release_assets: command not found
`. scripts/ci/gitea-release.sh` sources from the CHECKED-OUT TREE, and a release
rebuild checks out the OLD TAG. So the step could only ever see the helpers that
existed when that tag was cut, and the prune is gated on exactly that path: the
helper was guaranteed absent in the only case that calls it. Adding it to a shared
script made it look available at review time while being unreachable at run time.
Only the workflow file is read from the dispatched ref, so the logic moves there,
inline. Same reasoning documented at both ends, including the corollary worth knowing
before the next rebuild: a PKGBUILD fix made after a tag does NOT reach a rebuild of
that tag either — the packaging comes from the tag too.
Verified by executing the one-liner's exact bytes out of arch.yml under /bin/sh (the
shell Gitea actually uses): keeps the new -2 set and gamescope, drops the superseded
-1 packages and their .sha256 sidecars, leaves other legs' .dmg/.deb untouched. The
`'\n'` survives the shell quoting, which was the part worth proving.
Also drops the now-dead helper from gitea-release.sh rather than leaving a function
no caller can reach, and leaves a warning there against the next one.
Arch moved FFmpeg 8 -> 9 (every libav soname +1) hours before the release. PR #108
fixed the real bug — packaging/arch/PKGBUILD now binds punktfunk-host to the sonames
it actually linked, so pacman refuses an upgrade instead of bricking the install — and
re-keyed ci/arch-ci.Dockerfile so the builder would carry FFmpeg 9.
The tag was pushed four minutes later. arch.yml and docker.yml have no `needs:` between
them, and arch.yml deliberately runs no -Syu ("the image's snapshot IS the build
environment"), so the release build pulled the still-FFmpeg-8 `:latest` and published
punktfunk-host 0.25.0-1 depends: libavcodec.so=62-64, libavutil.so=60-64,
libavfilter.so=11-64, libavdevice.so=62-64,
libswscale.so=9-64
against a world that had moved to 63/61/12/63/10. It fails safely — pacman refuses,
nothing bricks — but it fails broadly: pacman prepares one transaction, so an
unsatisfiable dependency of OURS stopped affected users' entire `pacman -Syu`.
Nothing in the pipeline could have caught it. The existing assert proves the dep is
VERSIONED; it cannot prove the version EXISTS. So two guards, plus the lever to repair
a release that has already shipped:
* Preflight parity — compare the builder's libav `provides` against the live repos and
`-Syu` the container if they differ. The image is a cache and may lag; on this one
axis it may not. Syncs into a throwaway --dbpath so the container never sits in the
partial-upgrade state a bare `pacman -Sy` leaves.
* Publish gate — resolve every built package with `pacman -U --print` against a
PRISTINE --dbpath. Empty db means "nothing is installed", so every dependency must
come from the repos exactly as on a user's box. Resolving against the builder's own
installed set is what would hide this: a stale ffmpeg satisfies a stale bound.
gamescope stays best-effort (dropped from the upload with a warning, never fatal).
* workflow_dispatch(release_tag, pkgrel) — a published release cannot be repaired by
re-running its tag: pkgrel would stay 1, which is invisible to a box that already
recorded the broken build, and the workflow file at the tag can never carry inputs
added after it. Dispatched from main it takes the WORKFLOW from main and the SOURCE
from the tag, publishes to the stable repo at a higher pkgrel, and replaces the
release-page assets (prune_release_assets: upsert replaces by NAME, and a rebuild's
filenames differ, so the superseded package would otherwise stay one click away).
Verified on a real ffmpeg-9 box (.21, CachyOS) rather than reasoned about: the gate
rejects the published 0.25.0-1 host with the user-visible error verbatim, and passes
client, web, scripting and gamescope — 0 false positives across all five artifacts.
The parity snippet reads today's `provides` correctly (`-Si --dbpath` on an empty db
works; pacman does not wrap fields when piped). Version logic exercised on all four
paths: rebuild -> 0.25.0-2 stable, tag push and canary unchanged, pkgrel=1 refused.
Ships as punktfunk-host 0.25.0-2. README gains the pacman error and what to do about
it; CHANGELOG says plainly that 0.25.0's Arch packages were wrong.
`plugin-kit-v0.3.2` failed at its very first real step:
error: Duplicate package path
at bun.lock:71:5
InvalidPackageKey: failed to parse lockfile: 'bun.lock'
warn: Ignoring lockfile
error: lockfile had changes, but lockfile is frozen
`@punktfunk/host` was listed TWICE, byte-identically, at lines 69 and 71. I
introduced it: the lock had exactly one entry before 10a0ef32 and two after.
Running `bun install` to add the biome devDependency duplicated the `file:../sdk`
entry — the same `file:`-dependency lock corruption already recorded against the
web workspace's overrides.
Nothing else in the lock is wrong, so this removes the duplicate entry rather than
regenerating (a regenerate risks reproducing it, since the `file:` dep is the
cause).
Verified with the exact commands the publish workflow runs, in order:
`bun install --frozen-lockfile --ignore-scripts` (the step that failed) now
succeeds, then the file:-dep repair, `bun run check`, `bun run typecheck`,
`bun test` 67/67, `bun run build` — all clean.
No source change; 0.3.2 is unpublished, so the tag moves to this commit.
ffmpeg-next 8.1.0 could not accept FFmpeg 9 at all: ffmpeg-sys-next's version probe
covered avcodec majors 56..62 (the range is exclusive of its end), so libavcodec 63 fell
outside what it knew how to bind. 9.0.0 widens that to 56..63, which is what actually
unblocks Arch. Bump both pins — the unconditional Linux dep and the optional Windows
amf-qsv one — and the lock with them.
No API drift to fix. The crate major is a CEILING, not a target: one source tree still
spans FFmpeg 7.x/libavcodec 61, 8.x/62 and 9.x/63 via per-version cfgs, and every wrapper
symbol the NVENC-libav, VAAPI and amf-qsv backends name survives 8.1.0 -> 9.0.0
unchanged. The three hand-written #[repr(C)] hwcontext mirrors are the parts no compiler
checks, so they were re-read against the real headers rather than trusted:
AVCUDADeviceContext and AVD3D11VAFramesContext are byte-identical across 7.1/8/9, and
AVD3D11VADeviceContext gained two trailing UINTs in 8 that 7.1 lacks — which is why that
mirror deliberately stops at the common prefix, and why its assertions now say what they
do and do not buy you. They pin our layout, not libav's; a green build is not evidence.
The CI image is the step that makes this reach users. arch.yml deliberately runs no -Syu
("the image's snapshot IS the build environment"), so the builder stayed frozen on ffmpeg
8 no matter what Arch shipped, and a canary built from that snapshot could not satisfy the
soname dep the PKGBUILD now derives. Re-keying ci/ rebuilds it against ffmpeg 9.
Ubuntu and Windows deliberately stay put: the noble .deb bundles its own FFmpeg 8 behind
an rpath and strips the libav sonames from its Depends, and Windows bundles BtbN DLLs into
the signed installer — neither is exposed to the break, BtbN publishes no FFmpeg 9 build,
and moving either would re-qualify an encode stack to buy nothing.
Verified end to end on 192.168.1.21 (CachyOS, system ffmpeg 2:9.0-5, RTX 5070 Ti): host
builds clean and links libavcodec.so.63/libavutil.so.61/libavfilter.so.12/libswscale.so.10
with no unresolved sonames; the ffmpeg-8 compat shim is gone and the service runs with
NRestarts=0 and answers 401 on :47990; pf-encode's 67 tests pass; and a live synthetic
encode drives real NVENC hardware through FFmpeg 9's libavcodec to a decodable 1080p HEVC
stream (180/180 frames, FEC loopback 0 mismatches) with libavcodec.so.63 and
libnvidia-encode both mapped into the encoding process.
`depends=('ffmpeg' ...)` carried no version bound, and pacman is the only one of our
packaging formats that does not derive dependencies from ELF DT_NEEDED — rpm
auto-generates `libavcodec.so.62()(64bit)`, dpkg-shlibdeps emits `libavcodec62`, nix
pins the closure. So when Arch shipped ffmpeg 2:9.0-5 on 2026-08-08 and every soname
moved (libavutil .60->.61, libavcodec .62->.63, libavfilter .11->.12, libavdevice
.62->.63, libswscale .9->.10), a plain `pacman -Syu` walked every Arch/CachyOS install
straight across the break. The result is not a crash we can log: the dynamic loader
cannot start the binary at all, so it is exit 127 *before* main() in a systemd restart
loop, and because punktfunk-web is a separate bun service with no libav linkage it keeps
serving happily while :47990 has nothing listening — which reads as "the mgmt API is
broken" rather than "the host is not running". `ldd /usr/bin/punktfunk-host | grep
"not found"` is the one-line diagnosis.
Depend on the sonames instead of the package. Arch's ffmpeg declares the matching
`provides=(libavcodec.so=63-64 ...)`, and makepkg rewrites each bare `libfoo.so` listed
in depends into `libfoo.so=<soname>-<arch>` by reading the built binary's DT_NEEDED, so
the bound tracks whatever FFmpeg the builder linked against with nothing to hand-maintain
across the next bump. pacman now refuses the ffmpeg upgrade rather than bricking the
install. A hand-written `ffmpeg<2:9` would have gone stale on the very next major; not
bundling FFmpeg the way the .deb does, because that exists only because Ubuntu 24.04 LTS
is frozen on 6.1 and can never satisfy the dep, while rolling Arch always ships a current
one.
Verified on a real ffmpeg-9 box (192.168.1.21): the built package records
libavcodec.so=63-64, libavutil.so=61-64, libavfilter.so=12-64, libavdevice.so=63-64 and
libswscale.so=10-64, exactly matching DT_NEEDED, with the two libs --as-needed drops left
bare and satisfied by any ffmpeg.
The new arch.yml step asserts that expansion actually happened. If it ever stops — Arch
dropping the soname provides, someone tidying the entries out of depends — the dep
silently degrades to an unversioned name that any ffmpeg satisfies, which is exactly the
state that caused this, and it is invisible in a green build until a box bricks weeks later.
Found on hardware by the GOG plugin's own parity gate, on a box with exactly one
GOG game installed:
HKLM\SOFTWARE\WOW6432Node\GOG.com\Games -> 1 subkey (IRON NEST ...)
host's built-in scanner: 1 entry
plugin: detect: absent, 0 games
parity FAILED - 1 missing, exit 1
`reg.exe` ALWAYS echoes the full hive name in its output rows, never the
abbreviation it was given: query `HKLM\SOFTWARE\...` and every line comes back
`HKEY_LOCAL_MACHINE\SOFTWARE\...`. regSubKeys built its match prefix from the
`HKLM\...` string it was handed, so no line ever matched and it returned `[]` —
on every machine, for every key, always. Measured verbatim on .173:
reg.exe: [HKEY_LOCAL_MACHINE\SOFTWARE\WOW6432Node\GOG.com\Games\2013434102]
regSubKeys: []
Its only consumer is the GOG plugin, so the symptom was "GOG reports no games
installed" rather than an error — the same shape as the SQLite reader in 0.3.1:
a total failure that every layer degrades into an empty library.
The contract was wrong too, and the hive bug hid it. regSubKeys returned whole
key PATHS while the GOG plugin uses each result as a bare NAME
(`const key = \`${GAMES_KEY}\\${id}\``, and the subkey name IS the product id
that becomes `external_id`). Even with the prefix fixed, paths would have
composed nonsense keys. It now returns names, which is what the sole consumer
and its own comment always assumed.
Parsing is split into an exported `parseRegSubKeys(stdout, key)` for the same
reason `parseRegQuery` is exported — this is a text format that breaks quietly,
and it had NO test coverage at all. Six added, using the verbatim .173 output:
names not paths, multiple subkeys, grandchildren ignored, the queried key is not
its own subkey, case-insensitivity, and empty/error input. Four of the six FAIL
against the old behaviour.
0.3.1 -> 0.3.2. Gates: biome clean, tsc clean, 67/67 tests, build clean.
The kit had NO biome config and no lint script, while every plugin repo that
consumes it has both. So its source quietly drifted — unused imports, unsorted
imports, formatting — with nothing to catch any of it. Running biome here for
the first time reported 20 findings across 8 files.
Adds `plugin-kit/biome.json` mirroring the plugin repos' (tab indent, double
quotes, recommended lint preset, organizeImports), a `check` script, and
`@biomejs/biome` pinned to the same `^2.5.2` the plugins pin — without that pin
`bunx biome` resolved 2.4.6, which rejects the 2.5 `rules.preset` key.
Two deliberate differences from the plugin repos' copy:
* no `vcs.useIgnoreFile` — those are standalone repos with a .gitignore beside
the config; plugin-kit is a directory inside this one, and biome errors with
"couldn't find an ignore file". The `files.includes` exclusions cover it.
* `!examples/**/dist` instead of `!ui/dist` — the kit has examples, not a UI.
`css.parser.tailwindDirectives` is carried over and is load-bearing: without it
biome cannot parse `@theme` in src/theme.css and reports three parse errors on
CSS that is perfectly valid Tailwind v4.
Everything here is formatter/import churn except two real findings, both fixed:
* `Layer` (library/define.ts) and `Cause` (sync-engine.ts) were imported and
never used;
* test/spike-httpapi.test.ts read `(reg?.body as …).ui.secret` one line after
`expect(reg).toBeDefined()`. The optional chain undoes the assertion: had
`reg` been undefined the `.ui` access would throw a TypeError instead of
failing the test readably. Now asserted to the type system too.
Wired into plugin-kit-publish.yml as a `Lint & format` step ahead of Typecheck,
so this cannot rot again.
Gates after: biome clean (42 files), tsc clean, 67/67 tests, build clean.
The playback process callback sized its writes from the mapped buffer's
capacity — PipeWire's quantum-limit, 8192 frames ≈ 170 ms — instead of
the graph's per-cycle ask (pw_buffer.requested). Every cycle therefore
queued up to 170 ms of PCM downstream of the ring, and, worse, taught
JitterPolicy that the device drains 170 ms per callback: the underrun
floor (want + one frame) rose above any depth the A/V sync loop may
request, so sync measured audio ~280 ms late and was forbidden — by its
own continuity rule — from draining it. The first on-glass run of the
latency overhaul showed exactly that: audio buffer 272 ms, a/v +284 ms,
stable.
Honor requested (capacity remains both the ceiling and the fallback for
requested == 0), and log requested-vs-capacity once per stream in the
shape of the host's per-capture-open quantum line, so the next on-glass
report can say which one is sizing the writes.
Needs libpipewire >= 0.3.49 (2022-03) for the requested field; every
ship target clears that.
Verified on .21: cargo clippy -p pf-client-core --all-targets -D
warnings clean, 167 tests pass, fmt clean.
The Windows host does not build:
error[E0425]: cannot find value `OFF_INPUT` in this scope
--> crates\pf-inject\src\inject\windows\dualshock4_windows.rs:65:48
error: could not compile `pf-inject` (lib) due to 1 previous error
`dualshock4_windows.rs` writes the neutral report straight to `OFF_INPUT` in its
bootstrap path — correctly, and exactly as the DualSense and Steam Deck backends
do: the devnode does not exist yet at that point, so there is no reader to race
and no seqlock to take. Its steady-state path already goes through
`publish_input`, which is the v2.3 seqlock.
But the import list only names `publish_input`. `steam_deck_windows.rs` imports
`OFF_INPUT` explicitly for the same bootstrap write; this one was missed when the
list was edited to add `publish_input`.
One word in a `use`. No behaviour.
WHY CI DID NOT CATCH IT: `pf-inject`'s Windows backends compile only for
`*-pc-windows-msvc`, and the crate is host-side, so the client Windows workflow
never touches it. A cargo check from a Mac cannot stand in either — pf-inject
pulls punktfunk-core and therefore ring, whose C build wants MSVC headers, so the
cross-check dies in cc-rs long before it reaches this file.
FOUND BY: running windows-host.yml's own build line on the CI runner (.133)
against the v0.25.0 release tree before tagging —
`cargo build --release -p punktfunk-host --features nvenc,amf-qsv,qsv`. It fails
at `pf-inject`, which is step 1 of the host job, so a v0.25.0 tag would have
produced no Windows host binary, no installer, and no host asset on the release.
main moved another 62 commits (a8a4b11f -> fca9f42c), taking 0.25.0 to 391 since
v0.24.0. Five PRs: decode aliasing (#102), A/V sync (#101), gyro correctness
(#99), web console sweep (#100), Apple ATS (#103).
THE CORRECTION THAT MATTERED. The notes carried "Audio that falls behind the
picture pulls itself back … Android was worst, with no correction at all",
describing the jitter ring's buffer-shedding as if it were sync. It never was.
The host has stamped `pts_ns` on every audio datagram since long before v0.24.0
and EVERY CLIENT DECODED IT AND NEVER READ IT — verified in the v0.24.0 tree
(`crates/punktfunk-host/src/native/audio.rs:162` stamps it; the client audio
paths ignore it). Lip-sync was an emergent property of buffer depth, and it got
WORSE as video got faster, which is why shaving milliseconds off the audio budget
had never helped. That bullet is rewritten to say what is actually true, and A/V
sync takes a TL;DR slot.
It displaces the settings-BOM bullet, which was the weakest of the six as a
HEADLINE: conditional (only if the file was ever saved by PowerShell), partly
duplicated by the Windows non-C: entry, and it survives verbatim in Fixed. A/V
sync affects every user, every session, every client, with sound on — and unlike
most of this release it shipped broken in EVERY release we have ever made.
GYRO NEEDS AN UPGRADE NOTE, so it got one. The pipeline was wrong end to end and
is now measured against a real controller, which MOVES AIM SENSITIVITY: a pad
presented as a DualShock 4 reported gyro 40x fast (host-side), and a PlayStation
pad on Android reported ~30% short (client-side). At 40x nobody could have
compensated — gyro aim was unusable, not miscalibrated — but the Android ~1.4x
change is exactly the size a real person tunes around, so `## Before you update`
names it specifically.
DELIBERATELY NOT PROMOTED. The decode-aliasing program (#102) reads like a
catastrophe — H.264 decoding into a surface it predicted from on 297 of every 300
access units of every stream we emit, on both rungs — but it NEVER SHIPPED:
`git ls-tree v0.24.0 crates/` has no pf-vkdecode/pf-dxvadec/pf-vaadec/pf-bitstream.
It is a ship-blocker that was cleared, and writing "your picture was subtly wrong"
would be false for every reader. It contributes one clause to the decode entry
(every path is now checked frame-by-frame against a reference decoder; Windows +
Intel AV1 routes through Direct3D) and a full section in the changelog. Same
reasoning already applied to #96 and the rav1d abort.
Changelog gains the A/V sync mechanism (including that video is the master and
continuity outranks sync — the ring refuses a sync request that would break audio
on a jittery link) and the aliasing section, with the four independent reasons
four gates missed it: a structurally-blind conformance vector, a test that had
encoded the bug AS CORRECT, a vacuous assertion that could not fail, and the fact
that it streamed clean on glass. gpu_parity is 11 legs, not the 9 an earlier note
claimed.
Verified after the merge: lock diff versions-only 35/35, `cargo metadata --locked`
resolves (39 members), `cargo fmt --all --check` clean in both workspaces, notes
body 0 internal-vocabulary hits, Play notes 497/500 by android.yml's own gate.
Wire 2, C ABI 17, no new capability bits in this range.
A doc paragraph in `pic_av1.rs` wrapped so that "first at frame / 6. Releasing…"
put `6.` at the start of a line. rustdoc reads that as an ordered-list item
starting at 6, which makes the following unindented `///` line a lazy
continuation — `clippy::doc_lazy_continuation`, denied by `-D warnings`.
Reflowed so the number cannot begin a line. Prose is byte-identical in content;
only the wrap points move. No code, no behaviour.
WHY THIS MATTERS FOR THE TAG. `pf-dxvadec` is Windows-only, and no Windows leg
runs on a push to main — so main being green proves nothing about this. The
failure surfaces for the first time in a release tag's fan-out, which is exactly
what happened to the FIRST v0.23.0 tag: it went red on Windows clippy for this
same lint, and the cure was a tag re-point.
Caught pre-tag by re-running the lazy-continuation scanner over the tree while
preparing v0.25.0 (0 hits before this commit's parent merged the new decode
crates, 1 after). Cannot be verified by compiling here — the crate does not build
on macOS — so the evidence is the scanner plus the lint's own rule, not a clippy
run.
Moving the management API onto Network.framework left one request per
connection, so a library grid paid a TLS handshake per poster where the pooled
URLSession had shared one. And the Apple client -- unlike Windows -- never
cached art at all, so it re-fetched every poster on every visit.
ArtCache: a size- and age-bounded blob cache in the CACHES directory (every byte
is re-derivable from the host, so the system is welcome to evict it). Keyed by
the SHA-256 of the absolute URL, so host-proxy paths and store CDN URLs share
one cache without colliding. Reads touch the entry, so eviction is by last USE,
not last write. Empty bodies and data: URLs are refused -- neither is worth a
file. Defaults: 128 MB, 30 days.
Connection pooling: MgmtConnectionPool keeps up to four keep-alive connections
per host and makes further callers wait rather than opening more, which is the
part that matters -- a grid can ask for dozens of posters at once. A connection
the host dropped since we last used it is indistinguishable from a live one
until we write, so a REUSED connection that fails is retried once on a fresh
one; a fresh failure is a real failure.
Keep-alive means a response can no longer be delimited by the peer hanging up,
so HTTPResponseParser.messageLength finds the end from the framing itself --
Content-Length or the chunked terminal chunk plus trailers. Getting that wrong
would truncate a response or bleed one into the next, silently, so it carries
the bulk of the new tests. A connection with bytes left over after a response is
dropped rather than reused: we never pipeline, so anything trailing means we are
out of sync.
LibraryView closes the loader's pooled connections on disappear instead of
leaving sockets open on a screen the user has left.
16 new tests: message framing (both encodings, partial reads, back-to-back
responses, close detection) and the cache (binary round trip, key separation,
refusals, expiry, LRU eviction).
The previous commit bought the library back on VPN/remote hosts by declaring
NSAllowsArbitraryLoads, which works but is blunt: it drops ATS for ALL of the
app's URLSession traffic, and the only other traffic is third-party cover-art
CDN fetches -- the one surface we never wanted to open. It cost the TLS-version
floor, forward secrecy, and the cleartext-HTTP block on URLs the host supplies
at runtime (custom entries and scanner plugins carry arbitrary ones).
So take the host out of the URL loading system instead. MgmtTransport speaks
HTTPS over Network.framework, which ATS does not govern, and states the trust
rule we actually mean in a verify block: the leaf must hash to the fingerprint
pinned during PIN pairing. That is the same rule punktfunk-core has always
applied on the QUIC stream plane -- which is exactly why streaming kept working
over Tailscale while the library did not.
With that, the ATS dict is gone and ATS is fully enforced again. Cover-art CDN
fetches keep ordinary URLSession with full system trust evaluation and no client
certificate. LibraryTLSDelegate is deleted; nothing pins through URLSession now.
Also here:
- HTTPResponse: just enough HTTP/1.1 to read one GET -- status, headers,
Content-Length and chunked framing (hyper streams the art proxy chunked). A
body shorter than Content-Length throws instead of returning partial JSON,
which would otherwise read as "this host has no games".
- LibraryError.pinMismatch, so a re-keyed host says "pair again" rather than
sending someone to debug their network.
- 403 joins 401 as "unauthorized": both are the host declining the certificate.
- baseURL brackets IPv6 literals; the old string interpolation did not.
- 11 tests covering the framings hyper emits and the failure modes that would
otherwise be silent.
Known trade-off: no connection reuse yet, so each poster costs its own
handshake where the pooled URLSession shared one. Fine on a LAN, worth revisiting
for large libraries over a high-latency link.
It led with `--mgmt-bind 127.0.0.1`, a rare cause, and never mentioned the one
fact that actually explains the symptom: the library rides the management API on
a different port (47990) than the QUIC stream plane (9777), so it can fail while
streaming to the same host works. Field triage of exactly that case spent hours
on the stream path before anyone questioned the port.
Leads with that now, and names browser-testing the port as the fast split between
"unreachable" and anything client-side.
The game library rides the management REST API over HTTPS (TCP 47990) through
URLSession, authenticated by mTLS and pinned by SHA-256 fingerprint. The app
declared no App Transport Security policy at all, so it ran under default ATS --
which exempts only "local" destinations (.local, unqualified names, RFC1918 and
link-local literals) and applies the full policy everywhere else. The host
certificate is self-signed (and carries no SubjectAltName), so it cannot satisfy
that policy: the library loaded at 192.168.x and failed at the TLS layer on any
other address.
Field-reported against a Tailscale host. 100.64.0.0/10 is CGNAT, not RFC1918, so
the library failed there while streaming to the very same address worked -- the
QUIC stream plane is raw UDP and never enters the URL loading system. A WireGuard
peer or a public-IP host breaks identically.
Declares NSAllowsArbitraryLoads, which must stay the dict's only key: iOS 10+ and
macOS 10.12+ ignore it whenever a more granular ATS key sits alongside it. Trust
is unchanged -- LibraryTLSDelegate still pins the host by fingerprint and hands
every other origin (cover-art CDNs) to full system trust evaluation.
Its rows still read "never frame-hash parity-checked: the rung exports a tiled
dmabuf with no CPU-readable image, so parity needs a readback path that does not
exist yet". That readback now exists, and all SEVEN legs came back bit-identical
to libavcodec on RDNA3: vendored H.264 250/250, our host's low-delay H.264
120/120, vendored H.265 250/250, host low-delay H.265 120/120, HEVC Main 10
50/50 as P010, vendored AV1 250/250 of 274 decoded, and host low-delay 4K
two-tile AV1 60/60.
The two arms collapse into one, because the thing that split them — AV1 having
evidence the other legs lacked — is gone. Every leg now has the same evidence.
It stays `verified = false`, and the note says why in the words the
unproven-rung test requires: it has NEVER run on a second vendor and has never
been soaked. That is a real limit rather than a formality — every other verified
pair in this table earned it on more than one part, and the D3D11VA AV1 row two
entries up is a rung that passed on one vendor's driver while failing on
another's.
The second reason is not about evidence at all, and it belongs in the record
rather than in a commit nobody reads later: flipping this flag is a ROUTING
change. `native_rung_admitted` is `verified || !below.verified`, so a verified
VAAPI outranks Vulkan Video on every Linux AMD and Intel client — the Steam Deck
included. The parity result justifies that change; it should still be made on
purpose, by someone who wants it, rather than arriving as a side effect of
writing down a test result.
Every other decode rung earns `verified` with frame-hash parity against
libavcodec. VAAPI could not: it hands out a DRM-PRIME dmabuf whose memory the
driver tiles, so nothing could read its decoded pixels back, and all four of its
legs sat at "never frame-hash parity-checked".
That was never bookkeeping. The D3D11VA AV1 rung decoded 250 frames, streamed
4K60 through a clean five-minute soak, and produced WRONG PIXELS for 186 of 250
frames on NVIDIA and 245 of 250 on Intel. It looked perfect on glass; only the
goldens caught it, and the same defect turned out to be in H.264 on two other
rungs. VAAPI was the one rung where that class of bug could still be sitting
with nothing able to see it.
It is not. Measured on .25 (Radeon 780M, RDNA3, radeonsi, Mesa 26.0.3, VA-API
1.23) on 2026-08-08, against the SAME golden files the Vulkan and D3D11VA rungs
are held to, read across the crate boundary rather than copied:
H.264 vendored vector 250/250 bit-identical (7 from the flush)
H.264 our host, low-delay 640x480 120/120 bit-identical (3 from the flush)
H.265 vendored vector 250/250 bit-identical (2 from the flush)
H.265 our host, low-delay 640x480 120/120 bit-identical (0 from the flush)
HEVC Main 10, P010 50/50 bit-identical (2 from the flush)
AV1 vendored vector 250/250 delivered of 274 decoded, and
display frame 0 byte-identical to
libavcodec's own PIXELS
AV1 our host, 4K two-tile 60/60 bit-identical
⚠ ONE vendor. AMD/radeonsi only; no Intel iHD box has run these legs.
The readback that made it possible:
* `pf-vaadec`'s `va` module gains `VAImage` and `VAImageFormat`, hand-declared
with every size and offset measured off libva 2.23.0's real headers by
`layout-probe.c` and pinned as compile-time assertions — the same discipline
the decode buffers already keep. The trap: `VAImage::width`/`height` are
16-bit, so `data_size` sits at 60 and not at the 64 counting 32-bit fields
gives, and every field after them is two bytes earlier than it looks.
* `pack_two_plane` is the pure geometry — the crop to the picture, the padding
columns dropped per row, and the chroma plane taken from the driver's OWN
`offsets[1]` rather than from `pitch * display_height`, which is the 1088-row
smear this program has already paid for once. It needs no device, so ten CPU
tests cover it on macOS and in the container.
* `video_vaapi_native::parity` drives the seven streams above through the
production entry point and hashes what the rung DELIVERS, in delivery order,
tail included — so the delivery path is under test as well as the decode, and
a frame's surface comes from its own release token rather than from an
inference about which pool entry holds which picture.
THE READBACK CANNOT REACH THE PRODUCTION PATH, and that is structural rather
than a promise. `vaDeriveImage`, `vaCreateImage`, `vaGetImage`, `vaMapBuffer`
and the rest are resolved by a `#[cfg(test)]` type that dlopens libva itself;
the production `Libva` gains no field; `sha2` is a dev dependency. A CPU test
scans this file's own source and fails if any of those symbols is dlsym'd
outside the harness, so a refactor cannot quietly undo it.
Derive is not guaranteed, so both routes are implemented and neither is
optional: `vaDeriveImage` first, `vaCreateImage` + `vaGetImage` as the fallback
(which also detiles), and if neither yields the pool's own fourcc the leg FAILS
naming what the driver gave it. There is no skip path — a parity test that
passes because it could not read anything is the failure mode this program has
been bitten by three times. Both answer on radeonsi, the first frame of every
leg is read through BOTH and they must agree, and `PF_VAAPI_READBACK=getimage`
reproduces the H.264 leg's 250/250 through the copying route alone, so the
fallback is exercised rather than merely written.
And it can fail — proven, not asserted. Planting the real geometry defect this
driver's layout makes visible (rows read contiguously, ignoring the 512-byte
pitch behind a 320-wide picture) fails at display frame 0 with the full
localisation: 68312 luma and 14998 chroma samples differing, max |delta| 255,
luma bounding box (0,1)..(319,239) — and with the goldens forced through one
route, 250/250 diverging with "suspect the readback geometry". `compare` and
`localise` also have CPU counterfactuals, and a hardware leg proves the readback
reads real and DISTINCT pixels and localises a one-byte flip to the exact pixel.
⚠ One thing the hardware legs do NOT cover, found by planting the other defect
and watching it do nothing: radeonsi's decode surfaces for every fixture here
have no VERTICAL padding — `offsets[1]` is exactly `pitch * height` — so the
chroma-plane trap is untested on this driver, and `pf-vaadec`'s
`reading_chroma_at_the_display_height_would_have_been_caught` is the only place
it is checked at all. `probe_this_machines_readback_routes` now prints the
derived layout and says which of the two it is, so the next driver answers for
itself instead of being assumed.
The Apple half of the A/V sync overhaul; the Rust half is 12a53183 and this
mirrors its policy rather than re-deriving one.
The host stamps `pts_ns` on every audio datagram and the client decoded it into
`AudioPCM` — and then never read it. Video's `pts_ns` is used end to end (the
end-to-end meter computes a true glass-to-glass `displayed + clockOffset − pts`
per presented frame), so audio free-ran at whatever depth its jitter ring
happened to reach, video was presented on an independent path, and nothing ever
compared them. The A/V offset was an accident of buffer depths: it moved
whenever the ring ratcheted under underrun pressure, and it got WORSE every time
video got faster, because a quicker decoder lowers the video leg and leaves
audio's exactly where it was.
Video is the master:
audio_e2e = (now + buffered_ahead + clock_offset) − pts_ns
av_offset = audio_e2e − video_e2e (> 0 ⇒ audio behind the picture)
`AvSync` smooths that with an EWMA, ignores what sits inside a deadband no
listener can detect, refuses the implausible outright rather than clamping it (a
wall-clock step must not steer the ring), and proposes a depth. Swift refuses
one thing Rust does not have to: the arithmetic itself. The Rust controller
works in i128, while Swift has no Int128 at this tools version, so the terms are
combined with overflow-REPORTING arithmetic instead of the `&-` the latency
meters use. That is not defensive padding — `ptsNs = 1 << 63` reads as
`Int64.min`, the difference lands on exactly `Int64.min`, and `abs()` of that
has no representable result, so checking the overflow flags AFTER the sanity
limit does not mis-measure the stream, it aborts the process from the audio
drain thread. The guard's short-circuit ordering is what makes the sanity check
safe to run at all.
Continuity outranks sync, always. `AudioRing.setSyncTarget` only ever takes a
REQUEST, clamped between the existing underrun-driven floor and the hard cap. A
link whose jitter genuinely needs more buffer than the picture is away keeps its
buffer and the residual is reported. `nil` is the default and reproduces the
previous behaviour exactly. The clamp raises its ceiling to the floor rather
than using it as-is: a device whose callback quantum alone exceeds the hard cap
makes floor > cap, and a plain `min(max(s, floor), cap)` would then hand back
the CAP — quietly below the continuity floor, inverting the exact ordering this
exists to guarantee, on the awkward hardware it exists to survive. (Rust's
`Ord::clamp` announces that condition by panicking; Swift would just get it
wrong, which is worse.)
The reference is the other half, and without it the loop is inert — which is why
this was split out rather than shipped alongside the Rust side. `LatencyMeter`
now publishes its most recent sample as a LEVEL, so the end-to-end meter the
presenter already writes per presented frame becomes the video figure the audio
plane reads. Both present paths (arrival and deadline) feed it without either
knowing audio exists, and the stage-1 fallback presenter — which stamps no
present at all — offers nothing, so the loop correctly declines to correct. The
level EXPIRES, unlike the Rust atomic: this client has a backgrounded keep-alive
that keeps audio playing and drops video decode entirely, and a reference with
no expiry would go on steering the ring against a figure minutes old and frozen.
And the reason none of this was visible: `bufferedMS`/`targetMS` existed only in
a periodic log line, absent from anything a surface could render. The HUD's
detailed tier now carries `audio buffer N ms · a/v ±N ms` and the 1 Hz stats log
gains the same pair, appended last so existing parsers are unaffected — both
numbers, because a deep ring on a jittery link is correct and only the offset
separates that from audio held late.
`PUNKTFUNK_NO_AV_SYNC=1` disarms the loop without a rebuild, as on the Rust
clients.
Verified: swift build + 225 tests (5 skipped) green. Every new gate was proven
non-vacuous by planting its own defect and confirming the gate caught it — the
continuity invariant, the clamp inversion, the deadband, both refusal paths, the
evidence threshold, the sync-pressure relax, the reference's staleness and its
survival of a drain, and `setSyncTarget` being live at all rather than dead
code, which is how the previous pass in this area shipped a correction that was
structurally unreachable with a green test. Two gates came back VACUOUS on the
first sweep and are the reason their inputs look so specific: the overflow test
was being caught by the sanity limit instead of the overflow guard, and the
refused-reference test was being caught by `latestSample`'s own `> 0` check
rather than by where the publish sits.
`finish` showed `outputs.last()` and retired every other picture an access unit
bumped out of the DPB without ever displaying it, and nothing flushed the DPB at
end of stream. Measured on .25 against the vendored vectors: 225 of 250 frames
for H.264, 204 of 250 for H.265, 45 of 50 for HEVC Main 10. D3D11VA and Vulkan
deliver every frame, so this was the rung's alone. All four legs now deliver
250 / 250 / 50 / 250.
The same function carried a second defect. `DmabufFrame::keyframe` was stamped
with the CURRENT access unit's `is_idr`, not the flag of the picture it was
about to display, and on a reordering stream those are different pictures: the
IDR is bumped out several units after it decodes and arrived flagged `false` on
all three legs' first frame, while a later AU draining the DPB flagged some old
trailing picture as a keyframe. That field is `DecodedImage::is_keyframe`, the
pump's post-loss re-anchor signal, so a mislabel re-anchors on the wrong frame.
Three changes, all inside this rung:
* **A deliverable queue**, the same shape as `video_vk_native`'s — extend, ship
the front, trim the oldest past the bound, count and rate-limit the drops into
`DecodeHealth::dropped`. Its DEPTH is derived differently and the divergence is
documented: the Vulkan rung's bound is `HOLD_HEADROOM - PIPELINE_HOLD` = 1
because a queued frame there counts against the pool ON TOP of the DPB's own
residency. Here the three claims are disjoint and a bumped picture MOVES from
`pending`/slot to `held`, so the queue inherits the claim rather than adding
one. The bound is the DPB's depth — the deepest carry-over a bump can leave —
and the measured cost is at most one surface (zero on H.264, whose three
seven-picture IDR drains are the deepest bursts these vectors have). A bound of
1 would have left 235 of 250 on H.264, most of the defect still in place.
* **An end-of-stream flush.** This rung has no EOS signal and cannot have one:
the pump feeds access units until the session ends and then drops the decoder.
So `flush` has the two honest callers — `Drop`, where nothing can be presented
and the job is to release the queue's surfaces and the DPB's before the pool
goes, and a caller that KNOWS the stream ended, which today is the conformance
harness. One walk, not a production path and an untested teardown path. AV1
needs none: it shows at most one frame per temporal unit and buffers nothing,
which its 250/250 says out loud.
* **`PictureFacts` recorded when a picture decodes**, and read back when it is
displayed. `keyframe` was the defect; `color` and `display` are the same
mistake one field along — an in-band HDR switch changes the VUI mid-stream and
AV1's render region is per-frame, so a queued frame shown two units later would
have been drawn with the newest picture's signalling.
Concealment answers `Ok(None)` and deliberately does NOT drain the queue, which
is the Vulkan rung's order and is load-bearing: `clears_demotion_streak` is
`delivered || !concealed`, so shipping a queued frame on a concealed AU would
zero the streak and take away the escape hatch that stops a rung concealing
forever from holding a frozen picture.
The three delivered-count assertions moved with the fix, and so did the CPU
derivation that reproduces them without a GPU — it now simulates the whole
delivery model (ledger, queue, one-per-AU hand-off, flush) in the order `decode`
does it, and carries the old behaviour beside the new one as a counterfactual:
a queue bound of 0 with no flush still reproduces 225/204/45 exactly, and the
test fails if it ever stops being SHORT. `settle` was split out as the pure half
of `finish` so the claim walk, the display ordering and the picture facts are
all assertable with no device; `the_queue_never_needs_a_surface_the_pool_does_not_have`
runs the surface-lifetime arithmetic over the real vectors and pins the peak
claims (9 of a 16-surface pool on H.264, 8 of 14 on both HEVC vectors), with an
unbounded queue as the counterfactual that shows the bound doing its job.
Gates run: `cargo fmt --all -- --check`, `cargo clippy -p pf-client-core
-p pf-vaadec --all-targets --features sdl3/build-from-source -- -D warnings`,
`cargo test -p pf-client-core --lib --features sdl3/build-from-source` (176
pass), the same filtered to `video_vaapi_native -- --include-ignored` (23 pass,
0 ignored) and `cargo test -p pf-vaadec` (48 pass) — all on .25 (Radeon 780M,
RDNA3, radeonsi, Mesa 26.0.3, VA-API 1.23); plus `cargo fmt --all -- --check`
and `cargo clippy --workspace --all-targets -- -D warnings` in pf-lxcheck2.
`ci.yml` runs `cargo clippy --workspace` on the HOST, where
`clients/android/native` and every `#[cfg(target_os = "android")]` module
elsewhere compile out, and `android.yml` only ever built. So the Android target
was never linted at all — not once. Five lints were sitting in
clients/android/native when this was noticed, in code no gate had ever read.
The gate is a Gradle task rather than a YAML step because cargo-ndk needs a
specific discovery environment (NDK sysroot, SDK cmake 3.22.1 for libopus,
`LIBOPUS_STATIC`, Ninja) and duplicating it into the workflow would let the lint
drift from the build — a lint that ran against a different toolchain is a lint
about a different program. `registerCargoNdkClippy` reuses the build task's
environment verbatim via the extracted `cargoNdkEnvironment`, so local and CI
runs are the same invocation.
It lints BOTH pointer widths, and that is load-bearing rather than thorough:
arm64-v8a is 64-bit and armeabi-v7a is 32-bit, so a cast that is redundant on
one can be required on the other. Linting only the primary ABI would license
"fixes" that break the 32-bit build — the shipping ABI for the many 32-bit
Google TV / Android TV boxes this client targets. x86_64 is skipped: it is
emulator-only and shares its width with arm64, so it costs lint time for no
signal the other two do not already carry.
The five resident lints:
* `audio.rs` / `mic.rs` `type_complexity` — the open-attempt closures now return
named `OpenedPlayback` / `OpenedCapture` aliases. The two tuples are mirror
images of each other (playback sends, capture receives), which the aliases now
say out loud.
* `vsync.rs` ×2 `unnecessary_cast` — **not** taken. `timespec`'s fields are
32-bit on armv7 and 64-bit on arm64, so the casts are REQUIRED on one shipping
ABI and redundant on the other; following the suggestion would break the
32-bit build. `i64::from`/`.into()` do not escape it either, they trade
`unnecessary_cast` for `useless_conversion` on the 64-bit side. Answered with
a documented `#[allow]` at the expression instead of in whichever build breaks
first.
* `pad_audio.rs` `needless_range_loop` — iterator form, preserving the
`channels < 2` no-op the range had.
Verified: `:kit:cargoNdkClippy` green on both ABIs, host-lane clippy for the
crate still clean, `cargo fmt --all --check` clean. The gate was proven
non-vacuous by planting `1i32 as i32` in an android-only module and confirming
it fails the task, then reverting.
The core, Linux, Windows and host halves of the audio latency overhaul landed
with Android deliberately left inert: `JitterPolicy`'s sync target defaults to
`None`, so this ring kept behaving exactly as it always had. What was missing
was not the loop but its REFERENCE — nothing here published where a frame
actually reached glass, and a controller with no reference is the mechanism you
can prove is present but that cannot act. This wires both halves.
The decode thread now reads the host capture `pts_ns` that every `AudioPacket`
has always carried and that this client, like every other, dropped on the floor.
Against the ring depth (published by the AAudio callback through the shared
`AudioSyncCell`) and the video plane's end-to-end figure it computes
audio_e2e = (now + buffered_ahead + clock_offset) − pts_ns
av_offset = audio_e2e − video_e2e (> 0 ⇒ audio behind the picture)
and asks the ring for a depth that closes it. Only ASKS: `set_sync_target` is
clamped between the underrun-driven adaptive floor and the hard cap, so a link
whose jitter genuinely needs more buffer than the picture is away keeps its
buffer and the residual is reported instead of being taken out of the listener's
stream. Continuity outranks sync, on this ring as on the others.
The reference comes from `DisplayTracker`'s `OnFrameRendered` callback — the one
place in the client that knows a frame truly latched — and it is computed ABOVE
the HUD gate now. A sync loop that only ran while the overlay was up would be
off on exactly the devices that report latency; the stats LOCK stays gated,
which is what that early-return was really protecting. Both decode loops feed
it, so sync works with "Low-latency mode" off as well.
Two deliberate refusals:
* The figure is published RAW. The HUD shaves the OS present floor off its shown
display/end-to-end numbers — metrics report what Punktfunk controls — but sound
has to reach the ear when the light reaches the eye, and a floor-shaved
reference would place audio a whole latch period early on every device.
* Below API 33 there is no render callback, so there is no confirmed present and
the loop stays inert (target `None` ⇒ today's behaviour exactly). The release
instant is NOT substituted for it: a release targets a FUTURE vsync and runs a
whole latch period (8-21 ms measured) ahead of glass, well outside the loop's
deadband — it would place audio early on every frame while looking like it was
working.
The plane is also no longer invisible. Ring depth and the smoothed offset ride
the stats array at 33/34 and the Detailed HUD carries `audio buffer N ms · a/v
±N ms`, the same wording the desktop HUD uses — both numbers, because a deep ring
on a jittery link is correct behaviour and only the offset separates that from
audio simply held late. The 1 Hz logcat line gains `av_ms` beside its depth, and
the depth itself now has ONE publisher: the counter copy is gone in favour of the
sync cell both readers already share.
The escape hatch is two levers. `PUNKTFUNK_NO_AV_SYNC=1` keeps the contract the
desktop clients document, but an app launched from the launcher inherits no
environment, so the one a field tester can actually reach is
`adb shell setprop debug.punktfunk.no_av_sync 1` — no rebuild, exactly like
`debug.punktfunk.presenter`. A loop that steers playback has to be bisectable on
the device that reports the regression.
Verified: `cargo ndk -t arm64-v8a check` clean; `cargo clippy -p
punktfunk-client-android --all-targets -- -D warnings` clean on the host lane CI
lints, and the Android target introduces no new findings (5 pre-existing lints in
audio/mic/pad_audio/vsync are unchanged — the android-gated modules are never
linted by the host workspace); `cargo fmt --all --check` clean;
`./gradlew :app:testDebugUnitTest` green. The new HUD test was proven
non-vacuous by planting the defect first — dropping the render call fails its
three positive assertions and leaves the three absence assertions passing, which
is the shape a test that "passes for the wrong reason" would not have.
design/audio-latency-overhaul.md W4. Apple (W6) still keeps today's behaviour.
The host stamps `pts_ns` on every audio datagram and the client decoded it
into `AudioPacket` — and then never read it. Video's `pts_ns` is used end to
end (the presenter computes a true glass-to-glass `displayed + clock_offset −
pts`), so audio free-ran at whatever depth its jitter ring happened to reach,
video was presented on an independent path, and nothing ever compared them.
The A/V offset was an accident of buffer depths: it moved whenever the ring
ratcheted under underrun pressure, and it got WORSE every time video got
faster, because a quicker decoder lowers the video leg and leaves audio's
exactly where it was. That is what a field report on the Steam Deck heard as
"the audio delay is way too high", and it is why shaving milliseconds off the
audio budget had not helped.
Video is the master. In a game streamer the video leg is the input-feel budget
and must never be inflated to satisfy the audio clock, while audio tolerates
small crossfaded corrections that are inaudible — and `crossfade_drop` already
applies them. So audio moves:
audio_e2e = (now + buffered_ahead + clock_offset) − pts_ns
av_offset = audio_e2e − video_e2e (> 0 ⇒ audio behind the picture)
`AvSync` smooths that with an EWMA, ignores what sits inside a deadband no
listener can detect, refuses the implausible outright rather than clamping it
(a wall-clock step must not steer the ring), and proposes a depth.
Continuity outranks sync, always. `JitterPolicy::set_sync_target` only ever
takes a REQUEST, clamped between the existing underrun-driven floor and the
hard cap. A link whose jitter genuinely needs more buffer than the picture is
away keeps its buffer and the residual is reported — sync can never starve the
ring into dropouts. `None` is the default and reproduces the previous behaviour
exactly, so the four client rings can adopt this one at a time without
diverging.
Two upstream defects found on the way, both prerequisites:
* The host stamped `pts_ns` at ENCODE time, inside the loop draining an
already-accumulated chunk, so every frame of a chunk carried near-identical
timestamps describing when we got round to encoding. Harmless while nothing
consumed it; a sync loop regulating against it would regulate against a
fiction. It now comes off the capture clock.
* The host did not pace. One capture callback hands over a whole quantum — 5 ms
when the graph honours our ask, 21.3 ms on a VM, where stock PipeWire raises
`min-quantum` to 1024 — and the loop drained all of it into back-to-back
`send_datagram` calls. The wire carried a 4-5 frame burst then ~21 ms of
nothing, and a ring can only absorb that by standing a burst period deep.
Frames now leave on the audio clock, which costs no average latency.
And the reason none of this was visible: `buffer_ms`/`target_ms` existed only
as a `tracing::debug!` line, absent from `Stats`. On a Deck the client runs
under Steam's `reaper` with stdout on a pipe nobody can read, so the one number
identifying a deep ring was unobtainable on the device reporting the latency.
The HUD now carries `audio buffer N ms · a/v ±N ms` — both, because a deep ring
on a jittery link is correct and only the offset separates that from audio held
late. The host also reports its negotiated quantum against the one it asked
for, per capture open rather than once per process.
Verified: 364 core + 40 presenter tests on Linux, clippy -D warnings clean on
punktfunk-{core,host} + pf-{client-core,presenter}, fmt clean. New tests pin
the safety invariant (sync cannot pull the target below the continuity floor on
any preset), that `None` leaves the policy bit-identical, and that a device
quantum exceeding the hard cap does not panic `Ord::clamp` inside a realtime
callback.
Android and Apple keep today's behaviour (the `None` default) until their
presenters publish a video figure to align against; design/audio-latency-
overhaul.md carries the plan.
The H.264 and H.265 rows still read "NEVER decoded a frame on any hardware".
That stopped being true on 2026-08-07, in the same session that proved AV1:
every access unit of the vendored H.264 (250), H.265 (250) and HEVC Main 10
(50) vectors was accepted on .25 (Radeon 780M, RDNA3, Mesa 26.0.3) with no
decode error — NV12 for the 8-bit legs, P010 for Main 10, all on the same tiled
AMD modifier — and probe_this_machines_libva reports VLD decode for all three
profiles.
The row records the delivered counts honestly rather than rounding them up:
225/204/45 against 250/250/50 access units, because `finish` shows
`outputs.last()` and drops the other pictures an AU bumps, and nothing flushes
the DPB at end of stream. That is this rung's own behaviour — D3D11VA delivers
all 250 — and it is invisible on punktfunk's zero-reorder host output. It is
recorded and asserted rather than fixed: changing the one-frame-per-AU contract
touches the pump's deliverable queue, an end-of-stream flush, and the
`keyframe`-labels-the-access-unit defect in the same function, so it belongs in
a commit that moves all three.
Still `verified = false` for all four, and the note says why in the words the
unproven-rung test requires: never frame-hash parity-checked. That is not
pedantry — the D3D11VA AV1 row two lines above is a rung that decoded 250
frames and produced wrong pixels for every one of them. Parity is what
distinguishes them, and this rung exports a tiled dmabuf with no CPU-readable
image, so it needs a readback path nothing has written yet.
Every AV1 frame either decode rung has ever been measured against is `tile_cols =
tile_rows = 1`. The vendored vector is single-tile on all 274 of its frames, so every
tile array the conversions fill — `tiles.widths`, `tiles.heights`, the per-tile records
— had only ever been written at index 0, and a conversion that wrote tile 0 and left
the rest zero would pass the whole suite. Our encoder splits 4K into TWO TILE ROWS.
**The fixture.** `lowdelay-3840x2160.ivf.av1`, 261 KB, 60 frames — `punktfunk-host
spike --source synthetic --codec av1 --width 3840 --height 2160 --fps 60 --seconds 1
--bitrate 1` on .21 (NVENC, RTX 5070 Ti), wrapped to IVF with `ffmpeg -f obu … -c copy`
so `common::split_av1_aus` (the vendored parser's own `IvfIterator`) frames it exactly
as it frames the vector, with no second splitter that could disagree.
**4K is not a size choice, it is the only shape with the property.** Measured on the
same box with the same command: 1280x720, 1920x1080 and 2560x1440 all give `tile_cols =
tile_rows = 1`; 3840x2160 gives `tile_cols = 1, tile_rows = 2` with
`width_in_sbs_minus_1 = [59]`, `height_in_sbs_minus_1 = [16, 16]`, and both tiles in ONE
Tile Group OBU. 60 frames instead of 120 pays for the resolution: 261 KB, under both the
282 KB H.264 and 270 KB H.265 low-delay fixtures.
Goldens are libavcodec's software decode, cross-checked between ffmpeg n8.1.2 (Arch
x86_64, libdav1d) and 8.1.1 (Homebrew, macOS arm64, libdav1d) whose 746,496,000-byte raw
outputs are BYTE-IDENTICAL, not merely equal per frame. 60 of 60 digests distinct.
**AV1's frame accounting is asserted, never derived.** The vendored vector is 250
temporal units carrying 274 coded frames of which 24 are hidden; this stream is 60 units,
60 coded, 60 shown, 0 hidden, 0 `show_existing_frame`, 1 key frame. Neither is the
general case, so both parity harnesses now take units / decoded / shown as three
independent parameters instead of computing one from another, and the CPU guard states
all six numbers.
**A CPU gate that needed no hardware at all.** `pic_av1`'s new
`a_two_tile_frame_fills_both_row_entries_and_leaves_the_rest_zero` pins the second row
entry against its OWN `height_in_sbs_minus_1`, requires the two rows to tile the frame
exactly, and requires TWO tile RECORDS out of ONE tile group with rows (0,0) and (1,0) —
the transposition a square grid could never reveal — each spanning real bytes. The
existing one-tile test asserts index 0 is right and `1..` are zero, which a broken
multi-tile conversion also satisfies.
⚠⚠ **This is a file, and on AV1 that distinction has already cost a release.** "250/250
delivered frames bit-identical to libavcodec" was true for the entire period the host was
shipping only the FIRST TILE of every 4K frame: the verification ran against a vendored
file while the truncation lived in packetisation, and the suite stayed green throughout.
This fixture closes the multi-tile gap on the DECODE rungs and closes nothing about
fragmentation, reassembly, loss or AU boundaries — the golden header, both module docs
and the leg docs all say so, at length, so the next reader does not inherit the same
false confidence.
Legs: `low_delay_host_av1_every_frame_hashes_bit_identical_to_libavcodec` on the Vulkan
rung (11 ignored legs now) and on the D3D11VA rung, plus two non-ignored CPU tests.
Verified: 11/11 Vulkan parity legs on .21 (RTX 5070 Ti, 610.57.04), the new one 60/60
bit-identical; workspace clippy `-D warnings` and `cargo fmt --all --check` clean on .21.
Follow-up to a85e8452, closing the three items that sweep flagged and left.
SIXTEEN BROWSER DIALOGS, GONE. Every destructive action in an otherwise fully
branded console handed off to `window.confirm` — a grey OS box with the page's
URL in it, no brand, no red on a delete, and untouchable by any story or
screenshot, which is part of why it survived this long.
They are replaced by one promise-based surface (components/dialogs.tsx) rather
than a dialog per call site. The native calls were EXPRESSIONS — `if
(!confirm(…)) return;` — threaded through mutation handlers; rewriting each into
"hold the pending action in state, render a dialog, run it from onConfirm" would
have put dialog machinery in every section file and turned each linear handler
inside out. Returning a promise keeps them the shape they already were, and it
is what let the navigation guard come along too: TanStack's `shouldBlockFn`
accepts `Promise<boolean>`. `beforeunload` necessarily stays native — a reload
is the browser's dialog to draw, and it will not wait on ours.
No warning copy was rewritten. Each message was SPLIT at its existing sentence
boundary: the question becomes the dialog's title, the consequence its body,
and "Continue?" is dropped where the affirmative button now carries the verb
("Delete", "Uninstall", "Unpair", "Stop every session"). 16 new keys, en and de
in parity at 629.
Verified by driving the real dialogs in a headless browser — all seven contract
checks pass, including the two that would be invisible until they bit: Escape
SETTLES the promise (an unsettled one would hang a mutation handler forever with
no error), and a cancelled prompt resolves null rather than "", so a caller can
still tell "backed out" from "cleared the field".
FOUR OF THE SEVEN NUMERIC FIELDS became InputNumber; three deliberately did not,
and now say why in place. The layout X/Y pair had a real defect: a screen left
of the origin has a negative coordinate, and `Number("-") || 0` rewrote the lone
minus sign to "0" before the digits could be typed. Measured on the built page:
the field can now be emptied to retype instead of snapping to its floor, and 900
in a 1..=16 field clamps to 16. The three left alone cannot take it — the grace
seconds field writes to the HOST on blur (InputNumber commits while typing, so
its clamp would race the apply), and the library's year/players are OPTIONAL,
where `value: number` has no way to say "unset" and would invent a year for
every entry without one.
The select's highlighted row moves off @unom/ui's neutral grey onto the brand
wash the nav and the preset cards already use.
The Displays story earned its keep immediately: adding `useDialogs` to that page
broke it in Storybook, because the provider was mounted in __root and nowhere
else. It belongs beside the other app-level providers in .storybook/preview.
The D3D11VA and Vulkan rungs both decoded into a surface they were predicting
from, on 117 of 120 access units of our own host's low-delay H.264 (`1c54d099`
for AV1, `834b2443` for H.264). `pf-vaadec` feeds `reference_frames` from the
same `plan.dpb_refs` snapshot, releases its whole `removed` list inline exactly
as the two broken conversions did, and neither fix commit touched it. It is
still exempt — this is the evidence, and the thing that keeps it true.
**Measured on the CPU, no GPU needed.** `walk_for_aliasing` drives the planner
and `plan_to_va` over both streams and counts four shapes. On
`lowdelay-640x480.h264` the aliasing PRECONDITION is fully present: 117 of 120
access units remove a picture their own `dpb_refs` still names, and on the same
117 the setup picture is handed the slot of a picture that access unit READS —
the D3D11VA/Vulkan defect verbatim, in this conversion, today. On the vendored
conformance vector both counts are 0, which is why that vector proved nothing
on two other backends for two milestones. Aliased submissions: **0 on both**.
**Why.** A slot is not a surface here. `plan_to_va` never invents one — every
reference it can name is read out of the `surfaces` table it is handed — and
the decode target is a separate parameter the caller takes from OUTSIDE that
table. `setup_surface` reaches the submission at exactly one field per codec
(H.264/H.265 `curr_pic.picture_id`, AV1 `current_frame` and
`current_display_picture`); HEVC is doubly safe, because its per-slice
`RefPicList` stores an INDEX into `reference_frames` rather than a surface.
AV1's documented substitution fallback is the one place the target can be named
as a reference, and only where the store resolved nothing at all to prefer.
**The exemption was incidental; it is structural now.** It needs the reference
table and the decode target to come from ONE snapshot of the bindings, and the
rung had that only by writing `free_surface()` and `surface_table()` adjacently
at three call sites. Split them and this rung acquires the defect exactly: the
table must be the PRE-removal one (that is where the references are), while a
free list consulted after the removals offers precisely the displaced picture's
surface. `Session::acquire_target` now returns the index, the surface and the
table together from `&self`, so a later edit cannot move one call and not the
other. No behaviour change: same order, same values, same refusal message.
Tests. `no_submission_names_its_decode_target_as_one_of_its_own_references`
(both streams, 0) with
`taking_the_decode_target_from_the_slot_table_aliases_on_the_low_delay_stream`
as the counterfactual that reproduces the defect on 117 of 120 — so the walk
demonstrably CAN see it when it is there.
`the_low_delay_stream_reassigns_slots_whose_pictures_it_still_reads` pins 0/250
and 117/120 so neither can drift silently.
`the_decode_target_can_never_be_a_surface_the_reference_table_names` sweeps
every binding state a 4-surface/3-slot pool can hold, and
`taking_the_free_surface_after_the_removals_would_hand_out_a_referenced_surface`
is the ordering counterfactual.
⚠ One existing test lost a VACUOUS half.
`the_setup_picture_routinely_inherits_a_just_freed_slot` asserted the decode
target was never also a reference while handing every picture its own
never-reused surface id — distinct integers cannot collide, so that assertion
could not fail whatever the conversion did. Its real measurement (225 of 250
access units reuse a just-freed slot, which is why the target is a parameter)
is kept; the collision half is gone, and the doc says where the question is
actually answered and why a recycling pool is what it takes to answer it.
Gates, run on `.25` (Radeon 780M, radeonsi, Mesa 26.0.3, VA-API 1.23), this
rung being Linux-only: `cargo fmt --all -- --check`; `cargo clippy -p
pf-client-core -p pf-vaadec --all-targets --features sdl3/build-from-source --
-D warnings`; `cargo test -p pf-client-core --lib --features
sdl3/build-from-source` (171 passed); the same filtered to `video_vaapi_native`
with `--include-ignored` (18 passed); `cargo test -p pf-vaadec` (48 passed).
Plus the pf-lxcheck2 container for the cross-platform half — fmt, clippy and
`cargo test -p pf-vaadec`, all clean.
All four VAAPI legs still decode with the refactor in place, not one access
unit refused: H.264 225 of 250 access units delivering a frame, H.265 204 of
250, HEVC Main 10 45 of 50 (P010), AV1 250 of 250 — the same counts and the
same tiled modifier 0x200000010401b04 those legs recorded before it. ⚠ The
H.26x legs live on `fix/vaapi-h264-h265-hardware-proof`, not on this branch, so
they were run by overlaying that commit's test module onto the scratch tree;
only the AV1 leg and the libva probe are reachable from here. This is a decode
measurement, not frame-hash parity — the rung exports a driver-tiled DRM-PRIME
dmabuf, so there is no CPU-readable image to hash. The alias assertions above
are the real evidence and they need no device.
⚠ NOT taken: `finish`'s `outputs.last()`, which ships one frame per access unit
and drops the rest of what a bump displaces (225/204/45 against 250/250/50),
with no end-of-stream flush. It cannot bite punktfunk — hosts emit zero-reorder
output, so `outputs` never holds more than one picture — and fixing it changes
`decode()`'s one-frame-per-access-unit contract with the pump (it wants a
deliverable queue, which `video_vk_native` already keeps) plus an end-of-stream
flush and the `keyframe`-labels-the-access-unit defect in the same function.
It is recorded and asserted on that other branch, whose three delivered-count
assertions any fix has to move in the same commit; doing that from here, blind
to them, would be worse than leaving it.
`fd6241a2` made HEVC's freedom from the release-ordering defect falsifiable on CPU and
recorded what was still missing: no low-delay HEVC stream was vendored, so the exemption
rested on a structural argument plus one throwaway measurement. This vendors the stream,
and the exemption HELD.
**The fixture.** `lowdelay-640x480.h265`, 270 KB, 120 pictures — `punktfunk-host spike
--source synthetic --codec h265 --width 640 --height 480 --fps 60 --seconds 2 --bitrate 1`
on .21 (NVENC, RTX 5070 Ti, driver 610.57.04). Deliberately the H.264 sibling's resolution
and frame count: the two are then directly comparable, 640 and 480 are both multiples of
MinCbSizeY so there is no conformance window and a hash mismatch can only be decode rather
than readback geometry, and 270 KB sits alongside the 282 KB already accepted for H.264.
Goldens are libavcodec's software decode, cross-checked BIT-IDENTICAL across ffmpeg n8.1.2
(Arch, x86_64) and 8.1.1 (Homebrew, macOS arm64), 120 of 120 digests distinct.
**The exemption held, measured rather than argued.** `sps_max_dec_pic_buffering_minus1 = 4`
against the four pictures 8.3.2 keeps marked in steady state, `sps_max_num_reorder_pics = 0`,
`numRefL0 = 1` — a five-picture DPB filled exactly by four references plus the current
picture. 115 of the 120 access units retire a picture, and `removed ∩ dpb_refs` is **0 of
120**. A 300-picture 1080p stream from the same host reports the same shape: 295
retirements, 0 intersections. It is the encoder and not the resolution, exactly as for
H.264.
**A zero proves nothing on its own, so the fixture is pinned by its counterfactual.**
`test-25fps.h264` reported zero for two milestones while every stream we ship aliased on
99% of its frames. So the guarantee here is not "we looked and it was fine": hand
`plan_to_dxva_h265` the marked DPB as it stood BEFORE `decode_rps` — the mutation a
snapshot move would cause, reconstructed exactly as `dpb_refs(N-1) ∪ {stored(N-1)}` — and
the alias appears on **115 of 120** access units, driven through the real conversion rather
than through planner arithmetic. If a regeneration ever produced a stream that reordered,
or a DPB deeper than its reference count, that 115 collapses to 0 and the tests say so
instead of continuing to pass.
**The two rungs are exempt for different reasons, and the asymmetry is now a gate.** DXVA
binds the whole marked DPB — `RefPicList` is spec-defined that way, and an RFI long-term
anchor has to survive in it — so its exemption really is `H265Planner`'s snapshot ordering,
one call away from being untrue. `plan_to_vk_h265` never reads `dpb_refs` at all:
`pReferenceSlots` is the slots the operation uses, so it binds the current RPS sets, which
`decode_rps` itself derives and which therefore cannot name a picture that same RPS just
dropped. A new test feeds that conversion the identical widened snapshot and asserts
nothing changes, so a future change making the Vulkan rung bind the marked DPB — a
legitimate thing to want, since a *Foll* anchor invisible to the hardware is the RFI
failure shape — fails loudly instead of silently acquiring the defect.
What the Vulkan pixel leg adds is therefore NOT aliasing coverage, and its docs say so:
it is the first HEVC frame either rung has decoded from our own encoder, under a DPB that
retires and reissues a slot on 115 of 120 access units back to back, where the vendored
vector's reordering keeps that eviction slack.
Legs: `low_delay_host_h265_every_frame_hashes_bit_identical_to_libavcodec` on the Vulkan
rung (10 ignored legs now, up from 9) and on the D3D11VA rung, plus three non-ignored CPU
guards that run in ordinary CI.
Verified: 10/10 Vulkan parity legs on .21 (RTX 5070 Ti, 610.57.04), the new one 120/120
bit-identical; workspace clippy `-D warnings` and `cargo fmt --all --check` clean on .21.
A pre-release sweep of the management console for two things that no type check
and no diff can catch: primitives that were never @unom/ui's, and animation
that a nested motion parent quietly cancelled.
THE PRESET TILES ALL LANDED ON THE SAME FRAME. @unom/ui's <Section> sets
`delayChildren: stagger(...)`, so a page whose cards are direct descendants of
it staggers for free — which is why every page but one looked right. An
<AnimatedCard> is ALSO a motion element and sets no `delayChildren`, and the
Virtual displays preset tiles are cards nested INSIDE that page's config card,
so that card became their timing group. Measured in a headless browser: the
opacity spread between the first and last tile was 0.00 across the whole
animation (six tiles in lockstep), and is 0.98 now — a ~100 ms cascade matching
the rest of the console. The four hand-rolled copies of the stagger container
collapse into one `<Stagger>` that carries the explanation.
FIVE FILES IMPORTED THE WRONG BUTTON. `@unom/ui/button` exports both a plain
`Button` and the `AnimatedButton` that this console's wrapper re-exports under
the same name — so `import { Button } from "@unom/ui/button"` compiles, renders,
and silently opts out of the mount animation and the hover/tap response.
Displays, SessionGame, GPU, Update and PendingDevices had dead buttons sitting
next to live ones.
THREE PRIMITIVES HAD NO WRAPPER, SO NOBODY REACHED FOR THEM. @unom/ui ships
form/select, form/textarea and form/checkbox; components/ui did not, and the
gap was filled with browser-chrome `<select>`, `<textarea>` and
`<input type="checkbox">` in the add-hook modal and both library forms. Select
needs the same token correction Tabs needed — upstream `text-secondary` is a
text colour, but here `--secondary` is a SURFACE, so the trigger's chevron and
placeholder rendered at near-zero contrast on the card behind them.
The hook timeout also stops accepting a value the host rejects: `min`/`max` on
a controlled `<input type="number">` are decoration (no form validation ever
runs), so 900 went into a field capped at 600 and failed later, at run time.
@unom/ui's InputNumber clamps on blur and lets the field be empty while you
retype instead of snapping to the fallback.
Storybook gains the page that had no story at all — the console's largest
config surface, and the reason this shipped unseen. Its <Card> wrapper is load
bearing: it reproduces the motion nesting that IS the bug.
Four findings, all real.
**`SlotMap`'s own docs had become false.** "feed it every `DpbUpdate` in decode order
(via `Self::apply` or `plan_to_vk`, which applies internally)" — `plan_to_vk` no longer
applies internally, which is the entire point of the change, and `release`'s docs named
it as one of the two things that may free a slot. A reader following those docs would
build the next caller wrong in exactly the way this commit's parent fixed. Both now say
which conversions defer, which one does not, and why H.265 is the one that does not.
**The deferred release warned on a legitimate event.** `release_deferred` warned per id
when a deferred release found no slot — but a renegotiation replaces the whole
`Session`, and with it the slot map, INSIDE `plan`, while the planner's own drain
reports every drained picture in that same access unit's `removed`. Every one of those
ids then misses, and nothing is wrong. `debug!`, with the legitimate cause named so the
illegitimate one stays diagnosable.
**HEVC's exemption was asserted only in its consequence.** `the_current_picture_is_
named_by_curr_pic_and_never_aliases_a_reference` checked that no reference shares the
decode target's slot — which on the vendored vector holds whether or not the reasoning
behind it does. That is precisely how the H.264 leg passed for two milestones. The test
now also asserts the PLANNER property the exemption rests on (`removed ∩ dpb_refs = ∅`,
falsified by moving `dpb_snapshot()` above `decode_rps`), and records that the low-delay
measurement was 0 of 300 against H.264's 297 of 300 from the same host and the same run.
It also records what is still missing: no low-delay HEVC stream is vendored, so HEVC's
freedom is a re-derivable argument plus one measurement, not a standing hardware leg.
**Two stale cross-references.** Both AV1 conversions told the reader the H.264/H.265
zero was "measured on reordering vectors and not a proof" — the open question this
commit's parent closed. They now say what the answer was.
The AV1 review round flagged the H.264 leg as "plausibly the same defect, traced in
source, not reproduced" and deliberately did not touch it. It is reproduced now, and
it is worse than the AV1 one: it fires on 297 of 300 access units of every stream a
punktfunk host emits, at 720p, 1080p and 2160p alike, on BOTH the DXVA rung and the
Vulkan one.
**Decided on the CPU, no GPU needed.** `H264Planner` snapshots `dpb_refs` in
`begin_picture`, BEFORE `finish_picture` runs 8.2.5's marking and C.4.5.3's bump, so a
picture the sliding window unmarks and the bump then evicts lands in both `dpb_refs`
(which `RefFrameList` is built from) and `dpb.removed`. The conversion released the
whole `removed` list and then assigned the decode target a slot; `SlotMap::assign`
takes the lowest free slot, which is the one just vacated. `CurrPic = N` and
`RefFrameList[k] = N`, in one submission.
The two conditions have to coincide in ONE access unit, and low-delay H.264 is exactly
what makes them: `max_num_reorder_frames = 0` means the evicted picture has already
been output, which is what makes it evictable at all. NVENC seals it by writing
`max_num_ref_frames = 3` ALONGSIDE `max_dec_frame_buffering = 3` — a DPB exactly as
deep as its reference count — so the window unmarks the oldest reference in the very
unit whose bump drops it. The aliased picture is `ref_idx 2` of a three-entry
`num_ref_idx_l0_active` list: addressable by any macroblock, not a spare.
**Why two hardware-proven codecs and four GPUs never saw it.** `test-25fps.h264` is
level 1.3 with no VUI `bitstream_restriction`, so `dpb_limit` falls back to A.3.1's
level ceiling and gives a 7-frame DPB against 2 reference frames — the window unmarks
two units before the bump can evict — and it REORDERS, which keeps an unmarked picture
alive past the unit that unmarked it. Two independent reasons, both properties of that
vector rather than of H.264. It measured zero and passed 250/250 throughout.
`data/lowdelay-640x480.h264` is vendored to close exactly that: our own host's output,
120 pictures, goldens from libavcodec cross-checked bit-identical across two ffmpeg
builds on two architectures.
**The fix is the AV1 fix.** `DecodePlanDxva` and `DecodePlanVk` grow
`release_after_decode`, the conversions hand the removals back instead of applying
them, and the callers release them once the decode op is issued. It costs no slot the
map does not have: `SlotMap::new` allocates `max_dpb_frames + 1` and the DPB never
exceeds `max_dpb_frames`, so a free slot always exists with the whole `removed` list
still held — measured, peak 4 of 4 on the stream that defers on 117 of 120 units.
The Vulkan rung breaks on it in both DPB modes and neither loudly: DISTINCT hands the
aliased reference the same array layer the setup writes; COINCIDE clears
`slot_image[setup]` in the binding sync and the reference then resolves to no bound
image, dropping out of `pReferenceSlots` with a `trace!`. Its deferred release runs on
the FAILURE paths too — the fallible region's Result is held rather than `?`-ed,
because seven exits sat between the conversion and the release and each would have
leaked a slot.
`a_full_dpb_bump_reuses_the_slot_but_the_pool_model_binds_a_fresh_image` asserted the
aliasing as "the planner's normal behaviour": an authored depth-1 stream whose AU1
references the picture it evicts. It now asserts the opposite, which is the defect in
two lines.
New evidence, all of it runnable: the CPU proof pins BOTH numbers (0 on the vector,
117 of 120 on the low-delay stream) so neither can drift silently; the ledger-pressure
test measures the peak; and a low-delay parity leg is added to `pf-vkdecode`'s
`gpu_parity` and `pf-client-core`'s `video_d3d11_native::parity` so both rungs are held
to what they stream rather than only to what they conform to.
The `damaged` path has cleared `Session::held[setup_slot]` since M7, for a reason
that now applies to the failure path too: the slot map says the surface holds THIS
picture while the surface still carries whatever the previous occupant decoded, so
a later `show_existing_frame` naming it blits the old picture's pixels with the old
picture's geometry and colour. The failure path never reached that far before —
`decode_into`'s error returned straight out of `frame_av1` — and the previous commit
made it continue so the slot releases could run.
Five findings from the adversarial pass, all real.
**The deferral predicate was vacuous.** `plan.dpb.removed` is ALWAYS a subset of
`plan.dpb_refs`: `Av1Planner::plan_frame` snapshots `dpb_refs` before any mutation
and `refresh_slots` can only report a picture that was in `self.slots` at that
moment. So `filter(|id| dpb_refs.contains(id))` was a condition that is never
false, the eager-release loop beside it could never release anything, and the test
assertion "only a picture the submission points at earns the reprieve" could never
fire. Now: defer every removal, say why in terms of the planner, and assert the
PLANNER's property (`removed ⊆ dpb_refs`) — which is falsifiable, and whose failure
would mean the conversion is releasing a surface `ref_frame_map` points at.
**The failure-path claim was overstated.** Holding the decode's `Result` closes
this frame's leak, not the unit's: `decode_av1` returns on the first failing frame
and abandons the rest of the temporal unit's plans, so their removals are never
released. 24 of 250 units carry a second frame. Named rather than fixed — what to
do with the frames after a failure is the pump's question.
**⚠⚠ The H.264 leg plausibly has the same defect, and the comment this change added
said it could not.** `pic.rs` builds `RefFrameList` from `plan.dpb_refs`, and
`H264Planner` snapshots that in `begin_picture` — BEFORE 8.2.5 marking and the DPB
bump. The vendored bump drops a picture the sliding window just unmarked once it
has been output, so a picture can land in both `RefFrameList` and `dpb.removed`:
the AV1 aliasing shape exactly. Measured zero on the vendored vector — but that
vector REORDERS, which is precisely what keeps an unmarked picture alive past the
AU that unmarked it. A punktfunk host emits LOW-DELAY H.264, where output happens
as each picture is decoded, which is the condition that makes eviction and
unmarking land in the same access unit. Traced end to end in source, not
reproduced (no low-delay vector). NOT fixed: changing a hardware-proven codec on an
unreproduced suspicion is the worse risk two commits before a release. Instead
`no_au_removes_a_picture_its_own_reference_list_names` makes the assumption
falsifiable, and its message says what to do when it fires. HEVC is structurally
safe and now says why: `H265Planner` snapshots `dpb_refs` AFTER `decode_rps`.
**Four more stale promotion sites**, past the four already fixed: `Backend::
NativeD3d11va`'s variant doc, `Decoder::new`'s Windows rung comment, `lib.rs`'s
module note and `clients/session/README.md`. Two sites that used the AV1 leg as
the live EXAMPLE of an unproven rung are marked as expired rather than deleted —
the reasoning is what the next bad-evidence leg will need.
**The AV1 dump was missing.** `PF_DXVA_DUMP` wrote h264 and hevc only, for the one
codec whose libavcodec capture has never been taken and where the dump is
therefore the only tool.
Reverts half of 1eab4b66 and closes G10's open frame question, both settled by the same
measurement.
1eab4b66 made two corrections to the Apple phone-gyro mirror. The negation was right and
stays: Apple reports the gravity VECTOR, pointing down, while an accelerometer measures
proper acceleration, pointing up at rest, and the wire carries the latter. The frame
change was wrong, and this removes it.
The mistake was a name collision. Two different frames are both called "the controller
frame". GCMotion reports a CONTROLLER in (Right, Forward, Up) — measured on a real
DualSense — which is not the wire's frame, which is why `GamepadCapture.forwardMotion`
converts. The mirror's orientation remap resolves THIS DEVICE into the frame its header
describes, x right, y up, z out of the screen. For the pose that mirror exists to serve —
a phone clipped upright with the screen facing the player — "out of the screen" points at
the player, so that frame is (Right, Up, Backward), which IS the wire's. It was already
correct. Applying the controller path's conversion on top rotated it out of true: a phone
sitting still would have reported gravity as −1 g on the roll axis rather than +1 g up,
i.e. claimed to be lying on its edge.
Reasoning by analogy is what produced it — "the mirror says controller frame, the capture
path says controller frame, so the same fix applies". Both files say it; they mean
different things.
What caught it was measuring the Android twin, which does the same thing straight through.
On glass: a DualSense on Bluetooth to a phone, streaming to a Linux host, reads +1 g on the
up axis end to end. Had the Apple mirror needed a conversion, the Android one would have
needed the same one and would have been visibly wrong. It is not.
The same run settles G10's frame, which shipped straight-through and explicitly unverified
because nobody had put a Bluetooth pad in front of the platform sensor framework. Now
somebody has. `PadSensors`' own first-sample log read `accel 0, 10000, 0` — exactly 1 g on
slot 1 — and at the far end hid-playstation published gravity as +0.991 g on ABS_Y, with
every rotation driving its correctly-named axis and the signs agreeing with gravity's
independent witness on 95 of 100 rotating samples. Android hands a controller's sensors
over in the pad's own frame, as documented. No remap, and the comment now says measured
instead of assumed.
Worth recording why the earlier suspicion was wrong, since it is the same trap in the other
direction: Android's DEVICE sensor frame really does put +z out of the screen, so a flat
phone puts gravity on z — but a CONTROLLER's sensors are reported in the controller's
frame, not the phone's. One platform, two conventions, chosen by what the sensor is
attached to.
Gate: Apple macOS `swift build` + full suite (215 tests, 5 skipped, 0 failures) and the
iOS-triple typecheck, which is what actually compiles `DeviceGyro.swift`; Android
`:kit:compileDebugKotlin`, `:kit:testDebugUnitTest`, `:app:compileDebugKotlin`. Green.
Still owed: `DeviceGyroRemap`'s four orientation matrices remain derived — this run used a
controller's own sensors, not the mirror, so it says nothing about them. They need a
gyro-less pad on wire index 0 and a phone turned through all four orientations.
`the_evidence_table_says_exactly_which_rungs_have_run_on_hardware` asserts the
same fact a third way — a proven list and a NOT-proven list, both spelled out —
so promoting the rung in the three places the handoff named still left a test
saying "the DXVA AV1 leg FAILS parity on two GPUs — claiming otherwise is the
dishonesty this program must not ship". It was right to fail; the pair moves
lists here.
Three prose sites that still described the leg as decoding wrong pixels move
with it: `native_supports_av1`'s device-facts note, `log_rung`'s honesty-surface
docs, and the OPEN question in the Windows Intel arm of `pick_native` — that last
one is marked CLOSED rather than deleted, because the question it raised (the
evidence filter asks "any evidence", and has no answer for BAD evidence) is a
real gap in the rule that outlived this particular leg.
The evidence table said these legs "have still never decoded a frame anywhere",
and VAAPI is the rung every Linux AMD/Intel client lands on. They have now
decoded, on `.25` (Radeon 780M / Phoenix1, RDNA3, radeonsi, Mesa 26.0.3, VA-API
1.23, /dev/dri/renderD128):
H.264 250/250 access units accepted, 225 frames delivered, NV12
H.265 250/250 accepted, 204 delivered, NV12
HEVC Main 10 50/50 accepted, 45 delivered, P010
(AV1, unchanged: 250/250 accepted, 250 delivered, NV12)
all on the same tiled AMD modifier (0x200000010401b04). Not one access unit of
any vector was refused.
Three `#[ignore]`d legs modelled on the AV1 one, plus the Annex-B access-unit
splitters they need — ported verbatim from `video_d3d11_native`'s test module so
the two platform rungs are driven over the same access units rather than over two
splitters free to disagree. Main 10 earns a third leg rather than a variation on
the second: ten bits is a different VAAPI profile, a different render-target
format and a different surface fourcc, and that leg's fourcc assertion is the
only thing that would catch a driver quietly handing back NV12 for a ten-bit
stream.
This is NOT frame-hash parity, and the doc comments say so rather than letting
the test names imply it. The Vulkan and D3D11VA legs hash every frame against
libavcodec because both can read their decoded surface back; this rung exports a
DRM-PRIME dmabuf whose memory the driver tiles, so there is no CPU-readable image
to hash without a `vaDeriveImage`/`vaGetImage` path production neither uses nor
wants. What these legs prove is that every access unit is accepted, that the
expected number of frames comes back, and that each one is a real exported
surface of the right shape and fourcc — enough to turn "never decoded a frame
anywhere" into a measurement, not enough to promote the rung to `verified`.
Two findings the run surfaced, neither of which bites punktfunk's own streams:
* The delivered counts are 225/204/45, not 250/250/50, and that is the RUNG, not
the driver. `finish` shows `outputs.last()` and never more, so an access unit
whose plan bumps several pictures out of the DPB displays the last and drops
the rest — 18 dropped at the H.264 vector's three draining IDRs, 45 on the
H.265 vector's 45 two-picture bumps — and there is no end-of-stream flush.
Hosts emit zero-reorder low-delay output with no B pictures, so `outputs` never
holds more than one picture in the field. A CPU-only test derives all three
counts from the planner alone, on any Linux box with no GPU, so they stay
explanations rather than recordings.
* `DmabufFrame::keyframe` labels the ACCESS UNIT, not the picture delivered:
`finish` is handed the current AU's `is_idr`. On a reordering stream the IDR is
bumped out several access units after it decoded and arrives flagged `false`,
while the access unit that drains the DPB at a later IDR flags whichever old
picture it displays as a keyframe. That flag is `DecodedImage::is_keyframe`,
the pump's post-loss re-anchor signal. Asserted so that fixing it is noticed,
not so that it is preserved.
Gates, all run on `.25` (this rung only compiles on Linux): `cargo fmt --all --
--check`; `cargo clippy -p pf-client-core --all-targets --features
sdl3/build-from-source -- -D warnings`; `cargo test -p pf-client-core --lib
--features sdl3/build-from-source` (169 passed); the same filtered to
video_vaapi_native with `--include-ignored` (16 passed). Plus the pf-lxcheck2
container's workspace-wide `cargo fmt --all -- --check` and `cargo clippy
--workspace --all-targets -- -D warnings`, both clean.
The evidence table in `video.rs` still says these legs have never decoded a
frame. It is being edited concurrently, so its replacement row is handed over
rather than raced for here.
Two halves.
**The harness.** `libav_picparams_parity` covered H.264 and HEVC only, which is
exactly the gap that let a wrong AV1 submission ship. It now plans, converts and
packs all 274 frames of the vendored AV1 vector and checks what needs no capture:
the three-buffer descriptor set with no quantization matrix (AV1's matrices are
selected by index, so `dxva2_av1_end_frame` passes NULL/0 and there is no buffer
to submit), no macroblock count anywhere, the 912-byte picture-parameter buffer,
and the tile records — which unlike H.264/HEVC slice records do NOT abut, because
a `DXVA_Tile_AV1` addresses a tile PAYLOAD and consecutive payloads are separated
by their `tile_size_minus_1` fields.
The one that matters most is `no_av1_submission_names_its_decode_surface_in_the_
reference_store`: the invariant the previous commit fixed, over the submitted
BYTES rather than over the plan. libavcodec cannot produce that shape — it fills
`RefFrameMapTextureIndex` from the pre-refresh store and takes
`CurrPicTextureIndex` from a frame the reference update has not run on — which is
the argument for calling it a defect rather than a convention.
`AV1_FIELDS` reaches into the eight nested blocks (`tiles.widths`,
`segmentation.feature_data`, …) so a future capture reports a field and not "260
bytes of tiles differ"; `field_table!` grew nested-path support for it. The
`#[ignore]`d `our_av1_picture_parameters_match_libavcodecs` and the capture recipe
are in place, and `the_dump_and_the_parser_agree…` now self-compares AV1 too.
⚠ NO libavcodec AV1 capture was taken and the module docs say so rather than
leaving an absent result to be read as a pass: `.221` has no MSYS2, no gcc and no
make, so a patched FFmpeg there is a toolchain bring-up, not a build. Everything
this file claims about libavcodec's AV1 side is READ out of `dxva2_av1.c` (n8.1).
That reading did turn up one live divergence, recorded at `pic_av1.rs`'s
`pp.width` and deliberately NOT changed: libavcodec sends `avctx->width`, which is
FrameWidth (pre-superres), where this crate sends UpscaledWidth. The two are equal
whenever superres is off, which is every stream that exists here, so the 250/250
result says nothing either way and a blind change would be unmeasured.
**The promotion.** `(D3d11va, CODEC_AV1)` is `verified` — 250/250 delivered frames
bit-identical to libavcodec on an RTX 3500 Ada AND an Intel Arc. All three places
move together: the evidence arm, the module table and
`every_rung_runs_and_the_unproven_ones_are_named`, whose `unproven` array loses the
pair and whose proven list gains it.
⚠ This changes rung SELECTION, not just a label. `verified` is what lets `auto`
pick D3D11VA ahead of Vulkan Video, so Windows Intel and unknown-vendor boxes —
where the ladder is `native-d3d11va → native-vk → sw` — now decode AV1 on D3D11VA
where they previously fell to Vulkan. Taken deliberately: ~10x the Vulkan leg's
speed, and the parity that promoted it was measured on an Intel Arc, which is the
vendor family the change moves. Still no soak on the goldens, and the notes say so.
Also: `frame_av1` holds the decode's `Result` instead of `?`-ing it, so both slot
releases run on the failure path. `decode_av1` notes an error and keeps the
session rather than rebuilding the slot map, so an early return leaked a surface
per failed frame and hit `SlotError::Full` after nine.
AV1 applies `refresh_frame_flags` AFTER the frame is decoded (7.20), so a frame
that reads a reference slot and then overwrites it is the ORDINARY case, not an
exotic one: 268 of the vendored vector's 274 frames do it, first at frame 6.
`plan_to_dxva_av1` released every displaced picture inside the conversion — which
is what the H.264 and H.265 siblings do with their whole `removed` list — and then
assigned the decode target a slot. `SlotMap::assign` takes the lowest free slot,
and the lowest free slot is the one just vacated. So the submission said
`CurrPicTextureIndex = N` and `RefFrameMapTextureIndex[k] = N` in the same breath,
on 268 of 274 frames: decode into the surface you predict from.
Neither vendored H.264 nor H.265 vector ever produces that shape (measured: zero
on 250 AUs), which is why an eager release survived two hardware-proven codecs and
opened on the first AV1 frame past the key frame's neighbourhood. HEVC even has
the invariant under test already — `the_current_picture_is_named_by_curr_pic_and_
never_aliases_a_reference` — and AV1 had nothing.
The Vulkan rung already carries the fix; this is the same contract, and the DXVA
constraint is the STRICTER of the two: Vulkan binds only the references a frame
names, while `RefFrameMapTextureIndex` declares the whole store, so every picture
the store still names has to survive the conversion. `DecodePlanDxvaAv1` grows
`release_after_decode` and `frame_av1` applies it once the decode op is issued —
next to the `refresh_frame_flags == 0` release that already waits for the same
reason. Peak surfaces held goes 7 of the 9 the pool allocates, so the spare slot
`SlotMap::new` adds is doing exactly the job it exists for.
Measured on hardware before the fix: Intel Arc got 245 of 250 delivered frames
wrong — 47% of luma at the first bad frame, max |delta| 242, chroma wrong too, a
frame predicted from the wrong picture — and the only late frame it got right was
the one intra frame, which names no reference and so could not alias. That reads
as a `primary_ref_frame` defect and is not one: PRIMARY_REF_NONE and "has no
references to alias" are the same frames.
main went from 35ba64ca to a8a4b11f while this branch sat open — 190 more commits,
taking 0.25.0 to 327 since v0.24.0 and making it the largest release so far. That
scale is what forced the restructure.
THE SPLIT. Through v0.24.0 the engineering detail lived in an `## Under the hood
(for developers)` section at the bottom of the notes. It had grown to 21 dense
bullets sitting under the user-facing half — the exact burying the voice rules
exist to prevent, and it would only have got worse here. So:
* `CHANGELOG.md` at the repo root is now the technical half, newest release
first. It opens with a version table that lists every protocol number
INCLUDING the unchanged ones, because "did the driver protocol move?" is the
question an embedder most often needs answered and "no" is a real answer.
Then breaking changes, capability bits, wire planes, env vars, and the rest.
* `docs/releases/vX.Y.Z.md` keeps a short `## For developers` pointer and
otherwise contains no internal names at all.
* The link targets the file AT THE TAG, not at main. A release's notes are
frozen; a main link would silently start describing a later release.
* README.md and TEMPLATE.md now document this as the ritual rather than
leaving it a one-off, including a new rule 7 requiring the TL;DR.
The notes are SHORTER than before the merge — 83 lines against 108 — while
covering three times as much. That is the point.
TL;DR, six bullets: FFmpeg gone from the desktop clients; AV1 at 4K shipping half
of every frame; HDR/10-bit leaving half the encoder idle; Windows hosts minting
their own audio devices; the controller sweep; and settings silently resetting.
The last bullet points at `## Before you update` rather than restating it.
WHAT I DELIBERATELY DID NOT PROMOTE. #96 (HEVC DPB) and the rav1d half of #97 are
ship-blockers for the NEW decode stack, not live field bugs — verified, `git
ls-tree v0.24.0 -- crates/` has none of pf-vkdecode/pf-bitstream/pf-dxvadec/
pf-vaadec and `dpb_limit` did not exist at the tag. Nobody on a released build
has ever hit them. They are in Fixed and in the changelog, not the TL;DR, which
is reserved for things a reader is living with today. (A memory note claiming
shipped 0.24.x clients carry the HEVC bug was wrong and has been corrected — it
was about to drive a release decision.)
Also kept out of user-facing Fixed: the plugin-UI empty panel (fixes the origin
split from earlier in THIS release, so it folds into that change — but its 47993
firewall consequence IS in `## Before you update`, because an upgraded host keeps
a 47992-only rule and every plugin interface goes blank), and the pad-audio
WASAPI path fix (repairs a Windows build break in never-shipped code).
HONESTY CARRIED FORWARD rather than smoothed. The notes say plainly that the
Android overlay change did not make the stream faster, that nothing is
recoverable from the Windows non-C: settings loss, that Deck HDR still needs
Steam's own display setting, and that VB-CABLE should be left installed. The
changelog's verification section says the FFmpeg-deletion milestone never
executed on a GPU, pf-vaadec has never decoded a frame anywhere, openh264 has
never run on glass, and controller audio has never touched a real DualSense.
Play's "What's new" refreshed and re-fitted: swapped the TV-profiles line for the
safe-area/notch fix, which is visible to every modern phone user. First draft came
out at 525 chars; trimmed to 497/500, verified with android.yml's gate logic.
Re-verified after the merge (Cargo.lock conflicted — took main's and regenerated
the bump): lock diff versions-only 35/35 against origin/main, `cargo metadata
--locked` resolves (39 members; fec-rs, pf-driver-proto, usbip-sim and the newly
vendored cros-codecs keep their own versions), `cargo fmt --all --check` clean in
both workspaces, doc lazy-continuation scanner 0 hits over 579 files, notes body 0
internal-vocabulary hits above `## For developers`.
Wire 2. C ABI 14 -> 17 (15 rumble floor, 16 pad audio, 17 session end reason).
Driver protocol 6 and gamepad channel 3 untouched. host_caps is down to its last
free bit (0x80) and video_caps has been full since 0.23.0 — both now stated in the
changelog rather than left to be rediscovered.
The follow-up was framed as "build the frame-hash parity harness the D3D11VA AV1
rung is missing, then flip hardware_verified to true". Both halves were wrong.
The harness was never missing. `video_d3d11_native`'s `parity` module has carried
`av1_every_delivered_frame_hashes_bit_identical_to_libavcodec` since M7 wired the
rung — wired to the SAME libavcodec goldens the Vulkan AV1 leg passes against,
with the display-order model that handles the vector's 24 hidden frames, sitting
`#[ignore]`d beside the H.264/H.265/Main10 legs. It had simply never been run on a
device; .173 was powered off the day it was written. What the old evidence note
called a missing harness is real about pf-dxvadec the CRATE, which cannot host one
— it links no D3D11 — but the device half lives here and was already done.
Run on .221, it FAILS, on both GPUs, deterministically (three runs each, identical
first-divergent frame and identical hashes): 186/250 diverging display frames on an
RTX 3500 Ada, 245/250 on an Intel Arc.
It is the decode that is wrong, not the measurement, and three independent checks
say so. H.264 and H.265 pass 250/250 and HEVC Main 10 50/50 through the same
harness, the same readback geometry, the same crop and the same slot map on those
same two GPUs. pf-vkdecode's Vulkan AV1 leg reproduces the same golden file 250/250
on the same box. And the goldens regenerate byte-for-byte from the ffmpeg build
their own header names.
Two signatures, and they are not one defect wearing two faces. NVIDIA is bit-exact
for display frames 0..=63 and then loses ONE 16x24 luma block — 174 pixels, max
|delta| 8, chroma untouched — on the frame whose order_hint first reaches 64, after
which every remaining frame is downstream of it through prediction. The stream
parks the key frame (order_hint 0) in BWDREF and ALTREF2 for its whole length, so
64 is where the distance to it reaches the edge of what get_relative_dist can
represent at OrderHintBits = 7. Intel is structurally wrong from display frame 4 —
47% of luma, max |delta| 242, chroma wrong too, a frame predicted from the wrong
picture — and the only later frame it gets right is the one whose primary_ref_frame
is PRIMARY_REF_NONE.
None of this is visible on glass, which is the whole argument for goldens: the rung
streams 4K60 on both parts with a clean five-minute soak at roughly ten times the
Vulkan leg's decode time. The 2026-08-07 field sessions that looked clean were
looking at wrong pixels.
So hardware_verified stays false, and the note now says why in the strongest
available terms — it prints at warn on every session that lands here, and "decodes
AV1 to wrong pixels" is what a support engineer needs to read. The pair stays in
`every_rung_runs_and_the_unproven_ones_are_named`'s unproven array; its note still
contains NEVER, because the pair has never PASSED parity, which is now a measured
statement rather than an absence.
Left deliberately unchanged: `auto` on Windows can still reach this rung for AV1,
and on Intel it is the arm that fires, because that vendor advertises no SAMPLED
usage on any decode profile so zero-copy Vulkan Video cannot run there. Barring it
trades visibly-wrong AV1 for the software rung, which cannot keep up at 4K and is
itself unproven. Which way that trade goes is a product call, so it is recorded at
the admission site rather than made silently here.
`av1_divergence_map` is kept, cleaned up and documented: it is what turned "186
frames differ" into a lead — one line per display frame, its verdict beside the
plan facts that could explain it, and an opt-in raw-NV12 dump. At a frame where one
vendor hashes correctly, that vendor's bytes ARE libavcodec's bytes and so a valid
reference for the other's, which is how "how badly" was answered without new
goldens. The tool that would localise the rest does not exist: pf-dxvadec's
libav_picparams_parity covers H.264 and HEVC only, so the AV1 conversion has never
been compared against libavcodec at the picture-parameter level either. That is the
next step, not another session.
Also in this file, since it is the same table and the same day: the VAAPI rung's
AV1 leg has now decoded 250/250 of the vendored vector on RDNA3 and its arm is
split from the H.264/H.265 ones, which genuinely have still never decoded anything.
It is unverified for the same reason as ever — no parity — and the D3D11VA row
above is exactly why that distinction is worth keeping: a rung can decode 250
frames and still be wrong.
Follow-up to the G10 merge. The new sensor path was written against main, which does not
carry this branch's G8 work, so it forwarded motion unconditionally — the one thing G8
exists to stop.
`deviceMotion` checked `forwarding` and nothing else. A Bluetooth DualSense in a session
that resolved to an X-Box backend would stream ~200 Hz of samples the host parses and
discards, for the whole session, exactly as the USB capture path did before G8. Not a
regression against shipped behaviour — the path is new — but it would have shipped the
defect back into a client that had just been taught not to have it.
`Slot` now carries `motionReaches`, asked once at open off the kind that pad DECLARED, in
the same shape `ExternalPad` already used. Per pad, not per session: under Automatic the
handshake carries the active pad's kind, so a couch with an X-Box pad on slot 0 and a
DualSense on slot 1 must not have slot 1's working gyro suppressed by slot 0's answer.
The notice moved to where the truth is known. `openSlot` knows only what kind a pad
declared, not whether it physically has a gyro — that is discovered later, when
`PadSensors` finds a gyroscope and calls `setDeviceHasSensorMotion`. Raising it there is
the only placement that both tells a player whose gyro is being dropped and stays silent
for the pads that never had one.
Also unified the last duplicate scale in the module. G10 hoisted the wire units into
`Gamepad` and pointed `DeviceGyro` at them, but `DsDevice` kept its own `20L` / `10000L`
— and `Gamepad`'s new comment claims every sender goes through one place, which was not
yet true. Two copies of a unit constant in one module is precisely the defect this program
opened with (a DualShock 4 blob 40× hot because a second copy had drifted), so the claim
and the code now agree. `val` rather than `const val` only because widening to Long is not
a constant expression; Long is deliberate, since the calibration arithmetic overflows an
Int before it divides.
Proven non-vacuous rather than assumed: changing `Gamepad.MOTION_GYRO_LSB_PER_DEG_S` from
20 to 16 now fails four named cases across three classes —
`DsDeviceTest.calibrationRescalesRawCountsOntoTheWireUnits`,
`.theHostsOwnBlobIsAPassthrough`, `.parseStateAppliesTheCalibration` and
`DeviceGyroTest.wireUnitConstants`. Before this change `DsDevice` would not have noticed.
The gate itself has no test, for the reason the surrounding code already documents:
`GamepadRouter` needs Android plus a live JNI handle, there is no Robolectric in this
module, and a mock would test the mock. It is argued at the call sites instead.
Gate: `:kit:compileDebugKotlin`, `:kit:testDebugUnitTest`, `:app:compileDebugKotlin`,
`:app:testDebugUnitTest` — kit 75 / app 67, 0 failures, counts read out of the JUnit XML.
The merge reconciles: 62 on this branch, plus 6 from main's DeviceGyroTest, plus G10's 7.
Android had two motion sources and both of them are USB claims. DsCapture
takes a Sony pad's HID interface away from the kernel; Sc2Capture does the
same for a Steam Controller 2. Everything else — a DualSense, a DualShock 4,
a Switch Pro, an 8BitDo, paired over Bluetooth — arrives as an ordinary
InputDevice. Its buttons worked, its sticks worked, and its gyro was dead,
silently, with no log line and nothing in the UI to suggest the pad had a
sensor at all. That is not one controller, it is the whole class of
controllers people actually pair to a phone.
The platform has had the answer since Android 12: InputDevice.getSensorManager
hands back a SensorManager scoped to that one controller, carrying its
TYPE_GYROSCOPE and TYPE_ACCELEROMETER. PadSensors registers a listener per
forwarded pad that has a gyroscope and sends the samples on that pad's wire
index. Below API 31 it registers nothing and the pads behave exactly as they
did.
It is built on DeviceGyro's shape, because the phone mirror had already paid
for these lessons. One dedicated HandlerThread, never the main one. Batching
off (maxReportLatencyUs = 0) — batching would trade away precisely the latency
gyro aim exists to avoid. 200 Hz requested, which is also the ceiling the
framework grants an app without HIGH_SAMPLING_RATE_SENSORS, so asking for more
would only be capped. And a feed that lets go of a pad still alive parks its
rotation at zero first: the host holds motion as state and re-emits it in every
virtual-pad report, so an angular velocity left behind is a pad that rotates
forever.
Two writers on one pad's motion is the failure this program has spent the day
unpicking, so the coordination is explicit in three places. A USB capture wins:
DsCapture.startUsb already calls releaseDevice at claim time, that closes the
slot, and the close now also takes the sensor listeners off — the claim makes
the InputDevice vanish anyway, but going through the explicit teardown is what
makes the ordering deterministic instead of a race against the platform's own
removal callback. The phone-gyro mirror stands down: registering flips a bit
the router reports through padHasOwnMotion, which DeviceGyro re-reads on every
sample and answers with its own zero park. And a pad with an accelerometer but
no gyroscope is deliberately NOT taken — it could only send gravity while
pinning rotation at zero, on a pad the mirror is otherwise entitled to speak
for, which is the same fight in a quieter costume.
The wire units are measured fact (punktfunk_core::input::gamepad: 20 LSB/deg·s,
10000 LSB/g), and they now live in exactly one place on this client:
Gamepad.motionGyroWire / motionAccelWire, which DeviceGyro was hand-inlining a
second copy of. The gyro program's first finding was a client sending 40x hot
because a second copy of a number had drifted, and the merge that followed
found a sender nobody remembered to correct. One function, both callers.
THE AXIS FRAME ON THIS PATH IS NOT VERIFIED, and the mapping is deliberately
straight through rather than guessed at. What is known: the wire is a unit
passthrough into a virtual DualSense report, and that report's frame was
measured over raw HID on 2026-08-07 as (Right, Up, Backward-toward-the-player)
carrying (pitch, yaw, roll), right-handed — which is why the USB path forwards
the pad's own order un-remapped and is correct to. Android documents its sensor
frame for a handheld device as +x right, +y up, +z out of the face, the same
frame once "the face" is read as the one the player looks at. So straight
through is what the documentation implies. What nobody has done is put a
Bluetooth DualSense in front of the platform sensor framework and compare —
those numbers come through a HID driver and InputFlinger's sensor mapper,
either of which could permute or negate without saying so. A plausible-looking
wrong remap is exactly the bug this program keeps finding, so the code says
unverified and names the measurement that settles it, and each feed logs its
first converted sample so the cheapest half of that measurement — which slot
gravity lands on with the pad flat and still — costs a logcat line.
PadSensorsTest pins the scale, the clamp, the rounding and the straight-through
order, mutation-checked four ways: 20 to 16 fails gyroScaleFromRadiansPerSecond
and straightThroughFrame, reversing the axis order fails straightThroughFrame,
truncating instead of rounding fails roundsToNearestNotTowardZero, and negating
the accel fails restingPadIsTheHostNeutral. Its frame expectations are written
to change together with any remap that lands, not to be edited around one.
GamepadRouter needs Android and a live JNI handle and there is no Robolectric
here, so its half is argued in comments beside the code, as DsCapture's claim
ordering already is.
Gates: kit 65 tests (58 before, plus 7), app 67 unchanged, 0 failures, read out
of the JUnit XML rather than off a green build.
G17's motion half. The docs described what the CLIENT sends and stopped there, which
made a promise the host does not always keep.
The support matrix said a desktop client forwards motion from any pad SDL exposes a gyro
on "and the host injects it into the matching virtual pad". The first clause is true; the
second is only true when the virtual pad has a motion plane. The X-Box 360 and One
backends do not — no gyro in their HID contract — so the host parses every sample and
discards it. That is where *Automatic* lands anything it does not recognise as Sony or
Valve, an 8BitDo with a perfectly good gyro included, and where a Switch Pro lands on a
Windows host with no `hid-nintendo` backend to fold it into.
A reader following the old text would conclude their gyro was broken. The failure has no
other symptom: motion just does nothing.
So both pages now say what to do about it — pick a DualSense-class type — and the
client-settings page says it where the choice is actually made, next to the degrade
paragraph that explains why a session ends up on an X-Box pad in the first place.
The Deck's Steam-Input requirement moves out of Decky's settings blurb, which is the one
place a Deck user streaming FROM the Deck would never look. With Steam Input on, Steam
hands the app its own virtual X-Box pad, so no controller-type choice can help: there is
no gyro on the pad the client can see.
The picker help text now mentions motion on GTK and Android, which is where it was
missing — Windows already said it and Apple says it in its own words. One sentence, the
same sentence, so the four clients answer the question the same way.
This is the doc side of the on-screen notice that shipped earlier in this branch. The two
exist for the same reason and now agree: the client says it when it detects the case, the
docs say it when someone goes looking.
Not covered: the preset COUNTS in note 1 ("Android and the console home offer six …
Windows and Apple offer five") are still unverified against the four pickers, and the
Apple picker's missing Steam Deck entry is a code gap rather than a doc one. Both are
noted in the plan and left for their own change rather than guessed at here.
Gate: Linux CI image fmt + `clippy --locked --all-targets -D warnings` on
punktfunk-client-linux (the GTK string is compiled) plus the core crates and their tests;
Android `:app:compileDebugKotlin` + `:app:testDebugUnitTest`. Green.
The evidence table has said "native VAAPI: has never decoded a frame anywhere
(M6/M7)" since the rung was written. That is no longer true. Measured on `.25`
(Radeon 780M / Phoenix1 RDNA3, radeonsi, Mesa 26.0.3, VA-API 1.23, Ubuntu
26.04 — headless, no display server needed):
VAAPI AV1 rung constructed: native-vaapi av1
VAAPI AV1: 250 frames delivered, first 320x240 fourcc="NV12"
modifier=0x200000010401b04
250 of 250 displayed frames, first try, on the same vendored vector the Vulkan
and D3D11VA AV1 legs walk. The count matters as more than a smoke test: the
vector carries 274 coded frames in 250 temporal units — 24 units carry two, and
those extras are HIDDEN (decoded, referenced, never shown) — so 250 delivered is
this rung agreeing with the other two about which frames are output. A tiled AMD
DRM modifier rather than a linear one says the surface is a real decode target,
not a fallback.
Two changes, both in the rung's own file.
**The probe never asked about AV1.** `probe_this_machines_libva` walked H.264
High, HEVC Main and HEVC Main 10 and stopped there, which is part of why "never
decoded a frame" could stand so long without anyone noticing what had not been
asked. It now covers both AV1 profiles, and this box answers:
H.264 High: VLD decode AV1 Profile 0: VLD decode
HEVC Main: VLD decode AV1 Profile 1: no (VAProfile not supported)
Profile 1 being refused is correct — 4:4:4 AV1, which radeonsi does not do — and
it is the negative case that proves the probe reports rather than assumes.
**`av1_decodes_the_vendored_vector_on_this_machines_vaapi`** is the decode
itself, `#[ignore]`d beside the probe.
It is deliberately WEAKER than the Vulkan and D3D11VA AV1 legs, and the docs say
so rather than letting the name imply parity: those two hash every frame against
libavcodec's goldens because both can read their decoded surface back. This rung
hands out a DRM-PRIME dmabuf whose memory the driver tiles, so there is no
CPU-readable image to hash without adding a vaDeriveImage/vaGetImage path that
production neither uses nor wants. So it asserts what can be asserted honestly —
every temporal unit accepted, the right number of frames back, each a real
exported surface of the right shape, the first flagged as a keyframe — and it is
NOT frame-hash parity. Promoting this rung to `verified` still wants parity, and
parity wants a readback path first.
It fails loudly rather than skipping when the device has no AV1 entry point. It
is `#[ignore]`d, so it only runs when someone points it at a box that is supposed
to have one, and a silent pass there is exactly the invisible-failure mode this
program exists to end.
Gates: on `.25`, fmt clean, `clippy -p pf-client-core --all-targets -D warnings`
green under the Linux cfg where this rung actually compiles, the whole lib suite
167/167, and all 11 VAAPI tests green with `--include-ignored`. Workspace fmt +
clippy + lib suite also green in the Linux container.
⚠ Not touched here on purpose: the evidence table in `video.rs`. Its VAAPI row
still reads "never decoded a frame anywhere" and now understates what is known —
but a parallel agent is editing that same file for the D3D11VA AV1 row, so the
row is left for whoever lands second to update once, rather than conflicting.
Note for anyone reproducing on `.25`: it has no system SDL3 and no passwordless
sudo, so the test binary links only with `--features sdl3/build-from-source`
(SDL3 is gamepads, irrelevant to decode; production Linux still links the system
one). Its disk sits at ~99% full, and the tree there is a `git archive` export
with no `.git`, so `git apply`/`git checkout --` silently do nothing.
Surfaced by the merge. `DeviceGyro`'s header states the contract plainly — "units and
axis semantics match `GamepadCapture.forwardMotion` exactly … the same convention, so a
future sign/scale correction lands in one place for both sources" — and this branch made
two such corrections in only one of the two places. That is a promise the code stopped
keeping the moment the controller path was fixed.
Both were true parity when #88 was written; both broke here.
**The negation.** `GamepadCapture` sends `-(gravity + userAcceleration)` because Apple
reports the gravity VECTOR, pointing down, while an accelerometer measures proper
acceleration, pointing up at rest — and the wire carries the latter. The mirror sent it
un-negated, so a phone lying still told the host it was accelerating downward at 1 g.
The comment above that line even claimed the convention matched.
**The frame.** The mirror's remap targets the controller frame its own header describes —
x right, y up, z out of the screen — which is exactly GameController's frame, and that is
not the DualSense report frame the wire is defined in. So the same change of basis the
controller path now takes applies here, after the orientation remap rather than instead
of it: the remap resolves which way the phone is being held, and the basis change
translates the result into the pad's language. Two different jobs that happen to compose.
Order matters for the closing sample too. `stop` replays `lastAccel` beside a zero gyro
so "rotation stopped" does not also read as free fall; `lastAccel` is recorded after both
conversions, so what gets parked is what was actually sent.
Left alone deliberately: `DeviceGyroRemap` itself and `DeviceGyroRemapTests`. The
orientation matrices answer a different question — which way is the phone being held —
and nothing measured this evening bears on them. They remain derived-not-verified, as
their own doc says, and the on-glass pass that owes the controller path a check owes them
one too, in all four orientations.
Gate: macOS `swift build` + full suite (215 tests, 5 skipped, 0 failures) and the
iOS-triple typecheck green — the latter is what actually compiles this file, since the
whole thing is `#if os(iOS)`.
main moved ~60 commits while this branch was in progress, and one of them matters
here: PR #88 (the phone-gyro mirror) landed, touching the same motion path.
One conflicted file, `GamepadCapture.swift`, in three places — all of them the two
changes meeting rather than disagreeing:
- **Slot fields.** #88 added `motionSent` + `lastAccel` for its flush-parks-motion fix;
this branch removed `lastMotionNs` with the 4 ms drop-throttle. Kept both decisions:
the parking state stays, the throttle field goes.
- **forwardMotion's head.** #88 added the mirror stand-down (`pad 0` yields while the
phone speaks for it); this branch deleted the throttle guard. Kept the stand-down,
dropped the guard.
- **The send.** This branch converts into the DualSense report frame; #88 records what
went out so `flush` can replay it beside a zero gyro. Both, with the recording placed
AFTER the conversion — `flush` replays `lastAccel`, so it has to be the vector that
actually went on the wire, or a still pad's gravity gets parked in the wrong axis.
The two features compose exactly, which is worth stating because it is not luck: this
branch gates motion capture on `hasRotationRate`, and #88 engages the phone mirror when
`hasRotationRate != true`. They are complements — a pad either drives its own gyro or the
phone mirrors for it, never both and never neither.
Everything else auto-merged. Note `DeviceGyroRemapTests` is `#if os(iOS)`, so the macOS
suite reports the same 215 as before the merge rather than gaining #88's six — checked,
not assumed.
Gates re-run against the merged tree rather than trusting either side's: Linux fmt +
build + `clippy --locked --all-targets -D warnings` + punktfunk-core and pf-inject
suites; Apple 215 tests and the iOS-triple typecheck; Android kit + app compile and
tests. All green.
Re-measured against a host carrying #95, from .21 (RTX 5070 Ti, av1_nvenc) to
.221, on glass:
Intel Arc, auto -> native-d3d11va 4K60, decode 1.4 ms, e2e 16.7 ms p50
RTX 3500 Ada, pinned native-d3d11va 4K60, decode 1.0 ms
RTX 3500 Ada, pinned native-vulkan 4K60, decode 11.6-16.7 ms
Plus a 5-minute Arc soak: 297 stats lines, 60 fps, decode 1.3 ms, e2e 10.9/14.8 ms
p50, and exactly one WARN in the whole run — the hardware_verified=false notice
itself. No refusals, no demotions, no concealed runs.
Three things that follow.
The rung is no longer a one-session curiosity: it decodes 4K60 AV1 on TWO
vendors and survives a soak. The Arc leg matters twice over, because the Arc
advertises no SAMPLED usage on any decode profile — zero-copy Vulkan Video
cannot work there — so `auto` demoting to D3D11VA and then decoding is the
whole demotion path working as designed.
It is roughly 10x faster than the Vulkan AV1 leg on the SAME NVIDIA GPU. That
is the strongest argument yet for eventually letting `auto` pick it ahead of
Vulkan Video, which is exactly what `verified` gates.
And it stays `verified = false` anyway, because the missing piece is specific:
there is no frame-hash parity against libavcodec. Every other verified pair in
that table earned it with one, and pf-dxvadec has no harness that could produce
one — `libav_picparams_parity` compares picture parameters on the CPU and never
decodes a frame. Building that harness is the work that promotes this rung; a
fourth session is not. The evidence string now says so, so the next reader does
not have to rediscover which half is missing.
The VAAPI row is corrected in the same spirit rather than left as a bare "NO":
the reachable VAAPI box (.25, RDNA3) reports VAProfileAV1Profile0 /
VAEntrypointVLD and advertises no Vulkan AV1 decode at all, which makes it the
right box to prove that rung on and an unambiguous oracle when it happens. What
stopped it is recorded too — no punktfunk checkout there and 4 GB of usable RAM.
Documentation only — no behaviour change, and no flag flipped.
#95 disarmed sub-frame readback for AV1, which means AV1 forgoes the latency
win HEVC gets from shipping slice 1 while slice 2 encodes. The follow-up was to
teach the reader AV1's units: cut on OBU boundaries rather than byte counts and
arm from the driver's reported unit count. Measured on .21 (RTX 5070 Ti,
av1_nvenc) before writing any of it, and the measurement closes it rather than
scoping it.
Reading the frame headers av1_nvenc actually emits at 4K:
width_in_sbs_minus_1[0] = 59 one tile column, the full 3840
height_in_sbs_minus_1[0..1] = 16, 16 two tile rows
tile_start_and_end_present_flag = 0 BOTH TILES IN ONE TILE GROUP OBU
That last flag is the finding. "Cut on OBU boundaries" presumes the tiles are
separate OBUs and they are not — there is no boundary between them to cut on.
Shipping tile 1 early would need the HOST to re-author AV1 syntax per chunk,
synthesising a fresh Tile Group OBU header with tile_start_and_end_present_flag
= 1 and its own tg_start/tg_end. That is bitstream surgery on the encode path,
not the reader change it was assumed to be.
And the prize would be small even then, because split encode already spent it.
The two tile rows go to two split-encode engines that run CONCURRENTLY, so they
complete at nearly the same moment — the win is bounded by the skew between
engines, not by half a frame. Whole-frame encode measures 3.3-3.6 ms at 4K60
against a 16.7 ms p50 end-to-end, so even the sequential-tiles fantasy caps near
1.7 ms and the real number is a fraction of it. HEVC's win is bigger for a
structural reason that does not transfer: forced split and sub-frame are
mutually unsupported, so HEVC's slices genuinely are produced one after another.
1080p settles it further: tile_cols_log2 = tile_rows_log2 = 0, a single tile, so
there is nothing to pipeline at the commonest streaming resolution at all.
Recorded next to the disarm with the reopen condition named — NVENC emitting one
OBU per tile, or setting tile_start_and_end_present_flag = 1 — so this is closed
on evidence rather than left as an open maybe.
Documentation only — no behaviour change.
#97's frame-context floor closes the one rav1d abort we hit and can prove. It
does not make the rung panic-proof and nothing at that call site can, because
rav1d's public surface is dav1d's C ABI: any reachable panic crosses
`extern "C"` as `panic_cannot_unwind` and becomes `abort()`, past every
`catch_unwind`, rung demotion and typed refusal we have.
Counted across rav1d 1.1.0's 60 source files: 285 `unwrap()`, 214 `assert!`,
19 `unreachable!`, 11 `expect()`, 10 `panic!`. 539 sites that end the client if
a stream can reach them. #97 fixed one of them.
Process isolation is the only defence that actually works, and this records the
decision NOT to build it, with the reasoning, so it is not re-argued from
scratch each time someone reads that number:
* the defect is upstream's and is one line (memorysafety/rav1d#1497, filed
2026-08-07 with the fix and a reproducer; still open, no PR, as of today);
* 539 is an unbounded number, not a risk estimate — none of those sites is
known reachable from a punktfunk stream, and the honest next step is to
fuzz the rung and find out, which is cheap, rather than buy insurance,
which is not;
* the cost lands on the video path across Linux, Windows and Android (the
Apple clients decode through VideoToolbox and never reach this code), each
needing its own shared-memory frame transport, child lifecycle and
backpressure, and it adds a scheduling boundary to the slowest rung on the
ladder while zero-copy is a hard requirement;
* an abort here costs a session that was already degraded — this rung exists
because the GPU rungs failed first.
The trigger to revisit is named as an event rather than a feeling: a SECOND
distinct abort in the field, or a fuzzer finding a reachable panic. Either
makes it a class of bugs instead of one, and a class is what would justify the
architecture.
Documentation only — no behaviour change.
H.264 derives its DPB size the same way HEVC did before #96 — from a level
ceiling that says what a stream MAY use, not what it needs — and the ceiling
saturates at 16 frames, which is 17 hardware slots with the picture in flight.
That is the exact arithmetic that cost 720p and 1080p their HEVC.
Measured on real encoders (2026-08-07) rather than assumed: H.264 escapes it
twice over, and both escapes belong to the encoders, not to the format.
encoder level picked VUI restriction
NVENC (RTX 5070 Ti, 610.57.04) 3.2/4.2/5.1/5.2 present, buffering 3
VAAPI via libavcodec (RDNA3, 26.0.3) 4.1/4.2/5.1/5.2 present, buffering 1
openh264 (the software rung) 3.2/4.2/5.1/5.2 present, buffering 1
Every one picks a level proportionate to the picture AND states its real need
in the VUI bitstream restriction, so the ceiling is never reached and never
consulted. Nothing is broken today, and clamping would be wrong: with the
restriction present the number IS the stream's own statement, and a stream that
genuinely asked for a deep DPB would decode wrong if we shrank it.
So this does not change what any stream decodes. It gives the arithmetic one
named home (`dpb_limit`, the twin of `h265::dpb_limit`) carrying the evidence
and the reasoning, and it adds the signal that was missing: when an SPS carries
no restriction AND its level ceiling would demand more slots than mainstream
hardware provides, the plan now says so with `PlanWarning::LevelDerivedDpb`
instead of a user silently losing the codec the way #96's users silently lost
HEVC. It is not an integrity warning — the picture is intact; what fails is
opening a session — so `is_integrity_warning` classifies it false.
One thing the sweep corrects about how the follow-up was framed: it is SMALL
pictures that saturate the ceiling most easily, not 720p specifically. 640x360
at level 3.1 computes 16 as readily as 720p at level 5.0, because the ceiling
is MaxDpbMbs divided by the picture's macroblocks. The authored 64x64 test
fixtures land there too, which is why they now assert through `picture_warnings`.
Guards, as the missing consumer-end half of pf-encode's
`rfi_dpb_fits_a_mainstream_vulkan_decoder`:
* every_reachable_h264_stream_fits_a_mainstream_slot_pool — the measured
(picture, level, declaration) pairs, asserting slots <= 16
* the_level_ceiling_alone_would_reproduce_96_and_is_warned_about — the same
resolutions at levels that saturate, pinned WITH the warning
* a_proportionate_level_fits_even_without_a_vui_restriction — so neither
escape looks like it is doing all the work alone
Gates: fmt + clippy -D warnings clean; pf-client-core 167/167; pf-bitstream
84/84; and gpu_parity 8/8 bit-identical to libavcodec on the RTX 5070 Ti, which
is the gate that matters for anything touching the bitstream layer.
G13 — the three capture-fidelity findings from the gyro sweep, two fixed and one
argued.
**The 4 ms floor was a DROP, and it was shedding real rotation.** A sample arriving
3.9 ms after the last one was discarded outright. That is the wrong shape for this
signal: buttons and sticks are absolute state, so a dropped frame costs nothing — the
next one says everything it would have. Angular velocity is a RATE, and a consumer
integrates it into an angle, so a dropped sample is rotation that happened and can never
be recovered. GameController's delivery jitters around the pad's own ~250 Hz, so a floor
set AT that rate does not shed a rare extra sample; it sheds a steady fraction of every
turn. And the error is one-signed, so it accumulates — aim drifting short, which reads
as bad sensitivity rather than as a bug.
Nothing needed the ceiling. GC delivers at the sensor's rate rather than faster, the SDL
client has always forwarded every sample, and the host's idle watchdog is a 100 ms
timeout this cannot outpace. The throttle's two fields went with it: `lastMotionNs` was
left set-but-never-read once the guard was gone, and `motionIntervalNs` had no other
consumer. (Notes elsewhere say `flush` parks motion and reads it — that is PR #88's
branch, not this one. Checked rather than assumed.)
**An X-Box pad was streaming gyro it does not have.** Capture attached to any `GCMotion`,
and an X-Box controller exposes one that reports gravity and NOTHING else. So the client
sent a permanently-zero `rotationRate` to the host as authoritative gyro, under a
declaration saying this pad has one. That is worse than having no motion plane at all: a
game sees a controller being held perfectly still forever, and there is nothing to fall
back to and nothing to notice. Now gated on `hasRotationRate`, which is GameController's
own answer to the question we actually mean.
The settings badge had the same bug from the same cause — `hasMotion` was
`motion != nil`, so an X-Box pad got a gyroscope icon. It now reads `hasRotationRate`
too. One wrong predicate was driving both the UI promise and the wire behaviour, which is
why they were wrong together.
That also simplifies G8's "your gyro can't reach this session" notice, which had to test
`hasRotationRate` itself to avoid nagging about a gyro the pad never had. With the attach
gated on it, the notice is just the else-branch.
**Motion stays on the main queue, and this is the argument for why.** GameController's
`handlerQueue` is a property of the CONTROLLER, not of an element, so moving motion off
main moves buttons, sticks, the touchpad and the escape chord with it. This class is
`@MainActor` throughout — eight `assumeIsolated` sites, the slot table, the gesture
timers — so that is a rewrite of the isolation model rather than a queue assignment, and
it would put the tvOS escape chord (the only controller way out of a stream there) on a
background queue. That is a real risk for a speculative gain. The comment says so at the
call site, and names the measurement to make first if it ever does bite: the host's
per-pad motion inter-arrival histogram already reports exactly this and would say whether
the delay is client-side or on the wire.
Gate: macOS `swift build` + the full suite (215 tests, 5 skipped, 0 failures) and the
iOS-triple typecheck green. No test pins the throttle removal or the capability gate:
both are properties of live `GCMotion` delivery, which this module cannot fake — there is
no injectable seam, and inventing one to assert "we called sendMotion twice" would test
the mock. They are argued at the call sites instead, in the same spirit as the parts of
`DsCapture` that are not unit-testable in their module either. On-glass verification is
owed with the two already outstanding on that rig.
G14, unblocked by the frame measurement in efb7f991 — the plan deliberately left this
one alone until the up axis was known, on the grounds that a confidently wrong constant
would be worse than an obviously wrong zero. It is known now.
A virtual DualSense, DualShock 4 or Steam Deck that had received no motion reported
acceleration `[0, 0, 0]`. That is not "no data": zero proper acceleration means free
fall, which is a definite claim about the physical world and one that is never true of
a controller sitting on a desk or held in someone's hands — both read 1 g up. Anything
that interprets the accelerometer gets a confident wrong answer rather than a boring
right one.
It is worst exactly where it is least visible. A pad with no gyro at all — an X-Box
controller forwarded as a DualSense, which is what "Automatic" does for anything not
Sony or Valve — never sends motion, so it sits on that neutral for the entire session,
telling every game that reads it that the controller is falling. `switch_proto` has
always done this correctly on its own up axis, which is what made the gap visible in the
first place.
Which axis, and why it took a measurement. The wire is a unit passthrough into the
virtual pad's report, so the wire's up axis is the pad's own, and on 2026-08-07 a real
DualSense read over raw HID put `+0.997 g` on report axis 1 at rest, in a frame pinned
the same session as (Right, Up, Backward). So `MOTION_NEUTRAL_ACCEL` is `[0, 10000, 0]`
— NOT the z-up the notes had assumed from `switch_proto`'s documentation, which is why
guessing would have shipped a backend confidently disagreeing with the hardware.
The constant lives in punktfunk-core beside the units it is expressed in, and every
backend derives from it rather than restating it. The Deck's neutral in particular goes
through `steam_remap::motion_wire_to_deck`, the same rescale a real sample takes, so the
neutral and the live path can never end up with two opinions about what 1 g is — its
`hid-steam` resolution stays in exactly one place. The DS4 needs no separate change: it
reuses `DsState`.
`switch_proto` is deliberately NOT touched, and the test says so. It is a different
device on a different driver, its up axis is its own, and nobody has measured its frame
— aligning it to the DualSense for consistency would be the same unmeasured guess this
commit exists to avoid, just in the other direction.
Non-vacuity proven both ways rather than assumed. Moving the up axis to slot 2 (the old
z-up assumption) fails on the wire constant itself, which is what makes the measurement
load-bearing rather than decorative; reverting both neutrals to `[0, 0, 0]` fails on the
DualSense assertion with the message naming the defect. Each backend is checked in ITS
OWN units, because hard-coding "1 g" three times is how the halves of a unit contract
drift apart.
Gate (Linux CI image): fmt, build, `clippy --locked --all-targets -D warnings` across
punktfunk-core / pf-inject / pf-client-core, and both test suites — green, with
`Running tests/motion_contract.rs` and the new case's own `... ok` line observed in the
log rather than inferred from a green exit (`cargo test` stops after the first failing
binary, so a green-looking run can mean the contract test never executed at all).
G16 step 1, and the second half of what 9e9bb9f4 started. That commit fixed the SIGN
of acceleration (Apple reports the gravity vector, pointing down; a pad reports proper
acceleration, pointing up). This fixes the FRAME, which is a separate defect and was
never going to show up as an inverted axis — it shows up as roll where the game reads
yaw.
The wire is a unit passthrough. `dualsense_proto::write_report` puts gyro[0..3] and
accel[0..3] straight into the virtual pad's report bytes 16.. and 22.., in order, with
no permutation — the same slots a real DualSense fills. So the frame the wire is
DEFINED in is the pad's own report frame, and forwarding GameController's x/y/z
unconverted was speaking a different language with the same vocabulary.
Both frames measured 2026-08-07 from ONE physical DualSense on one desk, read twice —
over raw HID and through GameController — so this is two readings of the same
controller in the same orientations rather than two documents:
DualSense report frame: (Right, Up, Backward) axis 0 pitch, 1 yaw, 2 roll
GameController frame: (Right, Forward, Up)
Right is already slot 0; Up is GC's z and moves to slot 1; slot 2 wants Backward, which
is GC's y negated. Hence (x, z, -y), applied to gyro AND acceleration because it is a
change of basis and both live in that basis.
Notable: the wire's documented naming was right all along — gyro[0]=pitch, [1]=yaw,
[2]=roll is exactly what the hardware does. And Android needs no remap at all: it
forwards the pad's own axis order un-remapped, which is correct. Its old reading was
purely the scale bug f6de620f fixed. Only Apple was converting nothing.
How the hardware frame was established, since a wrong frame here is invisible. Gravity
at rest put +0.997 g on axis 1. Yaw clockwise-from-above drove axis 1 negative (98% of
the rotation), pitch nose-down drove axis 0 negative (100%), roll right-side-down drove
axis 2 negative (95%) — plain right-hand rule, and (a0 x a1 = a2) confirms the triad is
right-handed. The accelerometer then corroborated the gyro's assignment independently:
under pitch-down axis 2 rose 0.160 -> +0.339 (nose down raises the back, so world-up
gains a Backward component) and under roll-right-down axis 0 went +0.021 -> -0.197,
while yaw left acceleration untouched. Two different physical quantities agreeing on
one triad.
Apple's frame took four attempts, and the failures are worth recording because each was
a different way to be confidently wrong:
- peak |w| over a window containing BOTH the tip-down and the return stroke can record
the return, with the opposite sign. Yaw (a continuous one-way spin) was unaffected;
pitch and roll were exactly the two that disagreed with everything else.
- reading `gravity + userAcceleration` when `hasGravityAndUserAcceleration` is FALSE
yields a constant (0,0,1) in every orientation. It looks like data. The tell is that
it never moves. The client's own else-branch on `m.acceleration` is the correct read
and is what the instrument now mirrors.
- `da/dt = -w x a` holds only for gravity, so testing it during vigorous waving — when
`m.acceleration` carries inseparable linear acceleration — fits nothing.
The frame that survived all of that: static poses, three of them, three repetitions
each. Nose-down moved axis 1 by -0.635 (so axis 1 is Forward), right-side-down moved
axis 0 by -0.686 (so axis 0 is Right), flat put +0.99 on axis 2 (Up). That conclusion
holds whether or not the acceleration negation is right, because negating flips the
measured vector and the physical direction it represents together.
Confidence, stated honestly. The accelerometer half is solid: nine pose measurements,
and mapping the flat pose through gives (+0.005, +0.992, +0.192) against the hardware's
own (+0.021, +0.997, +0.160) — all three components, including the small tilt term that
is what distinguishes this mapping from the five other permutations that also put
gravity on slot 1. That the gyro shares the frame unmodified rests on a weaker
measurement: a gravity-dominated consistency test that preferred (+x,+y,+z) by 1.22x,
which is a margin, not a landslide. It is corroborated by the yaw reading (the one
rotation measured without the return-stroke ambiguity) agreeing with right-hand rule in
that frame, and by the peak-vs-return mechanism explaining the two that did not. A
device-side confirmation is still owed and is listed below.
The tests carry the measurements, not just the conclusion. Resting gravity is asserted
against BOTH readings of that pose; each rotation is asserted to reach the slot the wire
reads it from; and two properties guard the shape rather than the numbers — that the
conversion is an isometry (a basis change may not stretch anything) and that it
preserves handedness. That last one matters most: a permutation with the wrong number of
sign flips is a REFLECTION, which looks plausible axis by axis and inverts every
rotation. Mutation-checked: dropping only the negation fails 6 assertions across 4 of
the 5 cases, the handedness test among them.
Owed, and not claimed done: on-glass re-verification through a real iOS device, together
with the two already owed on that rig (the 9e9bb9f4 sign fix and the Android
calibration read) — one pass covers all three. G14's DualSense neutral acceleration is
now unblocked by this measurement (1 g on slot 1, not the z-up the notes assumed) but is
deliberately left to its own change; and that constant must NOT be propagated to
switch_proto, which is a different device whose frame nobody has measured.
Gate: macOS `swift build` + the full suite (215 tests, 5 skipped, 0 failures) with the
five new cases observed in the run's own output, and the iOS-triple typecheck green.
Two conflicts, both unions of independent removals/fixes: main fixed the
same three install.rs SAFETY comments this branch fixed (main's phrasing
kept), and the runner provisioning drops BOTH env lines — main removed
PF_FFVK_VULKAN_INCLUDE (pf-ffvk is gone since the FFmpeg replacement),
this branch removed VBCABLE_DIR (the retirement).
memorysafety/rav1d#1497, filed with the one-line fix and a reproducer that
needs no capture — any AV1 stream with one temporal unit removed. Written down
where the setting is, because the next person to read `av1_settings` and
wonder whether the floor is still needed should be able to check rather than
re-derive it.
`Set-Content -Encoding UTF8` writes a UTF-8 BOM, and every Windows how-to
reaches for it, so `%APPDATA%\punktfunk\client-windows-settings.json` edited
from a shell arrives with `EF BB BF` in front of the `{`. serde_json rejects
that at byte 0 — correctly, JSON has no BOM — and
`.and_then(|s| serde_json::from_str(&s).ok())` turned the refusal into
`Default`. Every setting in the file, gone, with the file plainly correct on
screen and not one word anywhere about why.
Cost an hour on 08-07: a `codec: "av1"` edit was ignored and the client
negotiated HEVC. The obvious suspects — the negotiation, the caps, the host —
were all working exactly as designed.
So the mark is stripped, which is what every other JSON consumer on Windows
does. But the BOM is only the instance; the bug is the `.ok()`, which hides a
trailing comma, a truncated write and a hand-edit typo just as completely.
Those now cost one `warn!` naming the file and serde's own line and column. A
file that cannot be READ at all is reported too, and for the same reason: PowerShell's
`-Encoding Unicode` writes UTF-16LE, `read_to_string` rejects it as invalid
UTF-8, and that lands in exactly the same hole.
The RESULT is deliberately unchanged — `Default`, never an error. Nothing about
streaming may hinge on a settings file being readable, and refusing to start
because one is malformed would be a worse failure than the one being fixed. A
missing file stays silent, because that is just first run.
All three of this client's JSON stores share the loader, because all three had
the identical line: the settings file, the known-hosts store (where a BOM
silently unpairs every host) and the profiles catalog.
The software rung aborted the process — not the session, the process — the
first time a 4K AV1 stream lost a frame. Reproduced on .21 twice on 08-07,
`SIGABRT` a few hundred milliseconds after "first frame decoded".
It was never about 4K, and it was never our bitstream.
rav1d 1.1.0 kills the process on ANY decode error while it holds a single
frame context. `rav1d_submit_frame`'s `c.fc.len() == 1` branch calls
`rav1d_decode_frame` inline; that always finishes in
`rav1d_decode_frame_exit`, which does an unconditional
`mem::take(&mut f.frame_hdr)` (decode.rs:4873); and then, only if the decode
returned `Err`, the same branch re-enters a local `on_error` whose first act is
`f.frame_hdr.as_ref().unwrap()` (decode.rs:4997) — on the `None` the teardown
just left. The panic unwinds into `dav1d_send_data`, which is `extern "C"`, so
it is `panic_cannot_unwind` → `abort()`: no `catch_unwind` at our call site, no
rung demotion and no `NoSoftwareRung` refusal can catch it. The same code is in
upstream `main` today, and 1.1.0 is the newest release, so there is no version
to bump to.
4K was only where an error first HAPPENED. The CPU rung cannot keep up at
3840x2160 (35-39 fps against a 60 fps stream), so the receive backlog stopped
draining, `pump::data` flushed it and jumped to live, and the next AU
referenced frames nobody had decoded. libdav1d gives the identical verdict on
the identical capture — 13 frames, then "Invalid data found when processing
input" — and simply carries on. At 1080p the rung keeps up, nothing is ever
flushed, no AU is ever damaged, and the same code ran for years without
anybody seeing this.
So the fix is to stop asking rav1d for the configuration whose error path is
broken. `c.fc.len() > 1` never calls `rav1d_decode_frame` at all: it hands the
frame to `rav1d_task_frame_init` and errors come back through `cached_error` /
`task_thread.retval` as ordinary `EINVAL`s, which the pump already answers with
a keyframe request. Measured, against the captured 4K stream:
n_threads=8 max_frame_delay=1 -> n_fc=1 -> ABORT
n_threads=1 max_frame_delay=1 -> n_fc=1 -> ABORT
n_threads=1 max_frame_delay=2 -> n_fc=1 -> ABORT <- proves the rule
n_threads=8 max_frame_delay=2 -> n_fc=2 -> 13 pictures, EINVAL, survives
n_threads=8 max_frame_delay=0 -> n_fc=3 -> survives
The third row is why `n_threads` grows a floor of two as well as the delay:
`n_fc` is `min(max_frame_delay, n_threads)`, so one decode thread silently puts
the whole thing back on the aborting path. That row is also what rules out the
theory this investigation started with — pinning threads to 1 was the suspected
trigger, and it makes things WORSE, so the tile workers are innocent and the
single frame context is the entire defect.
Two frame contexts would normally cost a frame of latency, and this does not,
because `decode` now drains PAST the first `EAGAIN`. `rav1d_get_picture` only
reaches its blocking `drain_picture` on a call whose own `drain` flag is already
set, and that flag is set by the PREVIOUS `get_picture` and cleared by every
`send_data` that carried bytes — so the first `EAGAIN` after a send does not
mean "no picture for this AU", it means "ask again", and this AU's frame comes
out of the second call. Stopping at the first `None` is what a
single-frame-context reading of dav1d's API teaches, and it would have put the
pipeline two frames behind while looking perfectly healthy. Measured over 14
temporal units at `n_fc = 2`: stopping at the first `None` produces nothing at
all for units 0 and 1; draining past it produces one frame per unit from unit 0,
at 20-42 ms per unit against `n_fc = 1`'s 21-53 ms. Not a trade — same cadence,
slightly faster, because the tile workers overlap the drain.
`Av1Software::new` then asks rav1d itself, through `dav1d_get_frame_delay`,
what those settings actually bought, and refuses to open a decoder that would
run with one frame context. That is not a restatement of the arithmetic: it is
`get_num_threads`' own answer, so it stays right if rav1d's derivation changes.
It is there because the failure it guards is uniquely quiet — an edit that
reinstates `n_fc = 1` costs nothing at build time, nothing in the tests and
nothing on a clean link, and then kills the client the first time a frame
arrives damaged. Losing the rung is recoverable; `abort()` is not.
On glass, .21, 35-second sessions, `PUNKTFUNK_DECODER=software`:
4K60 AV1 before: SIGABRT on the second frame, every run
after: exit 0, 0 panics, 35-39 fps, 1204 frames, decode_failed=0,
and 13 decode errors recovered from across 17 backlog
flushes — the exact condition that used to abort, survived
thirteen times in one session
1080p AV1 after: 40 fps, decode p50 2.2 ms (2.1 ms before the change)
What this does NOT buy: rav1d has other `unwrap()`s, and because its whole
public surface is dav1d's `extern "C"` ABI — every internal `rav1d_*` entry
point is `pub(crate)` — no in-process guard can turn one of them into anything
but an abort. This removes the one we hit and can prove; it does not make the
CPU rung panic-proof, and the evidence table says so.
Reported upstream with a self-contained reproducer: the in-tree
`test-25fps.ivf.av1` vector with one temporal unit dropped aborts rav1d at
`n_fc = 1`, survives at `n_fc = 2`, and libdav1d decodes it with 145 error
reports and no crash.
A punktfunk client streaming HEVC from .21 (RTX 5070 Ti) refused every access
unit with "stream needs 17 DPB slots, device caps at 16", flushed, waited for an
IRAP, got a fresh IDR that needed 17 too, exhausted the decode ladder and
reconnected with HEVC excluded. On a build with no software HEVC decoder — there
is no permissively licensed one — that is not a slower path, it is losing the
codec.
The host was blameless. Reading the SPS it actually emitted: general_level_idc
153 (L5.1 High, which NVENC autoselects at hevcConfig.level = 0 because a
130 Mbps target does not fit L5.0's 100 Mbps ceiling) and
sps_max_dec_pic_buffering_minus1 = 5 — six pictures, RFI_DPB references plus the
current one. Six, at every resolution. That is already the minimum the encoder
can honestly declare, and the only host-side lever, the level, cannot be lowered
without signalling a bitrate the stream exceeds. There was nothing to fix there.
dpb_limit was reading equation A-2 instead. A-2 is a CEILING on what an SPS may
signal — 7.4.3.2.1 constrains sps_max_dec_pic_buffering_minus1 to
0..=MaxDpbSize-1 — not a statement of what a stream needs, and it branches on
picture size against the LEVEL's MaxLumaPs. At 1080p the coded 1920x1088 =
2 088 960 luma samples fall under MaxLumaPs(L5.1) >> 2 = 2 228 224, taking the
first branch for min(4 * MaxDpbPicBuf, 16) = 16. max(A-2, buffering) then
reported 16 where the stream had asked for 6, the backends added one slot for the
picture in flight, and 17 is one more than NVIDIA's maxDpbSlots.
A resolution sweep on the box drew A-2's branch table exactly, and it is the two
commonest streaming resolutions that lost the codec:
720p 1280x720 = 921 600 branch 1 -> 16 frames, 17 slots 82 refusals, HEVC dropped
1080p 1920x1088 = 2 088 960 branch 1 -> 16 frames, 17 slots 41 refusals, HEVC dropped
1440p 2560x1440 = 3 686 400 branch 2 -> 12 frames, 13 slots clean
4K 3840x2176 = 8 355 840 else -> 6 frames, 7 slots clean, decode 1.9 ms
One host, one level, one six-picture requirement. Only which branch the picture
size landed in decided whether HEVC worked. That is also why this hid for so
long: 4K was the resolution it was exercised at, and 4K is the one size that
falls through to the honest answer. H.264 escaped for an unrelated reason — its
own level-derived ceiling happened to land at 13 for 1080p L5.0 and 5 for 4K
L5.2 — but it is the same shape of derivation and would fail the same way if
NVENC ever picked a higher level for a smaller picture.
So dpb_limit now returns the stream's own sps_max_dec_pic_buffering_minus1 + 1,
capped at 16. That is not a workaround, it is what the number means: it is
exactly the bound C.5.2.2's fullness clause bumps against, and A.4.1 bounds the
total RPS entries by the same value, so `buffering` pictures hold `buffering - 1`
references plus the current one with nothing left over.
The max() that produced the 16 was written to be generous to malformed streams —
"storing their pictures beats erroring the AU" — but it never did that either.
Dpb::needs_bumping (C.5.2.2) already keys on the signalled buffering, not on
max_num_pics, so a stream referencing more pictures than it declared was ALREADY
being bumped below its own declared depth before every store. The widened limit
bought no tolerance at all; all it ever did was over-allocate hardware surfaces,
by ten pictures per session at 1080p, and on NVIDIA take HEVC away entirely.
The fix moves 720p and 1080p onto the pool shape 4K has been running in the field
all along (7 slots, 6 references), so it is not a new operating point — it is the
one already proven. max_active_references drops from 15 to 6, still above the 5
an RFI_DPB stream can name. The per-AU level gate in pf-vkdecode reads
plan.picture.level_idc directly, so dropping A-2 out of NegotiationInfo costs no
sensitivity to a mid-stream level change.
Two regression tests pin the arithmetic from both ends, because either end
drifting back reproduces this:
- h265: the field SPS synthesized byte for byte on the fields that matter must
plan 6 frames / 7 slots, all four resolutions must agree because the stream
does, and every depth the envelope gate admits must leave room for the picture
in flight. The one honest residue is pinned too and deliberately left
refusing: A.4 does let a conforming stream declare a full 16-picture DPB, and
17 slots genuinely do not fit 16, so that stream is still refused rather than
decoded with too few slots and silently corrupted references.
- pf-encode: RFI_DPB + 2 <= 16, guarding the producer end. RFI is a real
latency win and this does not cap it at today's value — there are nine slots
of headroom — it just stops it being raised past the point where clients can
no longer decode us at all.
Two of its notes became false the moment the host stopped truncating AV1.
native D3D11VA / AV1 said "NEVER decoded a frame on any hardware". It has
now decoded 4K60 on an RTX 3500 Ada — and the same run is why the note
matters: its warn line named the rung as unproven moments before it
failed 72 access units running with "reference picture N holds no DPB
slot". That was the host shipping half of every frame, not the rung, so
the M7 wiring was right all along.
It stays UNVERIFIED regardless. `verified` gates `native_rung_admitted` —
whether `auto` may pick this rung ahead of Vulkan Video — and one
25-second session with no frame-hash parity and no soak does not buy
that. Promoting it wants a deliberate gpu_parity-style run. The note now
says what is true instead of what is convenient.
software / AV1 said rav1d had "CPU unit tests only". rav1d has now run on
glass: 1080p AV1 decodes, and 4K ABORTS THE PROCESS. It takes an internal
error path and panics inside its own on_error (rav1d 1.1.0
decode.rs:4997, unwrap on a None frame header); the panic crosses the
extern "C" boundary in dav1d_send_data, so it is panic_cannot_unwind and
no rung demotion or NoSoftwareRung refusal can catch it. libdav1d decodes
the same 4K stream 715/715, so this is rav1d's own defect and is recorded
where the next person to reach that rung will see it.
G8's Android half, and the last of the three clients. Same failure as the other
two: a controller with a gyro, in a session whose virtual pad has no motion
plane, does nothing when tilted — silently, with no way from the couch to tell
that apart from a broken sensor. The fix is the Controller type setting, so the
notice names it.
Android read neither the requested nor the resolved backend, so this needed a
plumb. What it did NOT need was a third copy of the rule. `nativePadMotionReaches`
takes the kind a pad declared and answers off `pad_motion_reaches` in
punktfunk-core, where the argument and the tests already live. The rule is
subtler than it looks — the host builds each pad from its OWN declaration and
folds what it cannot build, so neither the declaration nor the session echo
answers it alone — and every way of getting it wrong is silent. A Kotlin
transcription would have been a third thing to keep in step with the host, which
is exactly how the SDL half got it wrong the first time.
Asked once per pad, at claim, in `openExternal` — where the pad's kind is already
being declared to the host — and the answer held for the pad's lifetime on the
`ExternalPad`. Not per sample: this runs at a DualSense's full report rate.
`hasGyro` gates only the NOTICE, and defaults to false. `DsCapture` passes true —
every pad it captures is a Sony one whose IMU is a headline feature, forwarded on
the rich plane. `Sc2Capture` keeps the default, because the Steam Controller 2's
motion rides inside the opaque passthrough report that `hidReport` carries, which
nothing here may second-guess: warning about motion for a pad that never calls
`motion()` would be a notice about a feature the player never lost. The
suppression itself is on `motion()` regardless, where it costs a dead pad nothing
and stops a live one paying to send samples the host will decode and discard.
The notice sits at the BOTTOM of the stream overlay, unlike the mic-chord
confirmation at the top. The two can coincide — a pad is claimed at roughly the
moment someone might be muting — and one landing on the other would cost the user
both. It holds 6 s rather than the mic chord's 1.6: that one confirms something
the user just did, this one explains something they did not, in a sentence they
have to read. Nulled at teardown beside `onExitArmed`/`onMicChord`, for the same
reason those are — a slot closing during release must not poke Compose state on
the way out.
Not covered by tests, and this is a limit of the module rather than a choice:
`GamepadRouter` needs Android plus a live JNI handle, there is no Robolectric
here, and the predicate it defers to is pure Rust that already has its table. So
the parts that carry the reasoning are argued in comments, as `DsCapture`'s
claim/teardown ordering already is. What IS mechanically verified is the piece
that a compiler cannot catch and a device would fail on: the JNI symbol
`Java_io_unom_punktfunk_kit_NativeBridge_nativePadMotionReaches` is present and
global in the built arm64-v8a `.so`, so the `external fun` resolves rather than
throwing `UnsatisfiedLinkError` at the first pad.
Gate: `:kit:compileDebugKotlin`, `:kit:testDebugUnitTest` (62 cases, 0 failed,
read out of the JUnit XML rather than inferred from a green build — unchanged
from this branch's previous count), `:app:compileDebugKotlin` and
`:app:testDebugUnitTest` (67 cases, 0 failed), with `:kit:cargoNdkRelease`
rebuilding the JNI crate clean across all three ABIs, plus `cargo fmt --check` on
it. On-glass verification is owed on the rig the earlier legs used, and is worth
doing as one pass with the two already owed there.
G8's Apple half — the UI hint 77797a9e left owed, plus the suppression, which on
this client is worth more than it was on the SDL one.
The failure being fixed is entirely silent. A controller with a gyro, in a session
whose virtual pad has no motion plane, simply does nothing when tilted: nothing
in the app says so, and from the couch a session that resolved an X-Box backend
is indistinguishable from a broken sensor. The fix is the Controller type setting,
so the hint has to name it — a badge that only said "motion unavailable" would
leave the player exactly as stuck.
Asked per pad, off what the slot declared, via the predicate punktfunk-core now
carries. `GamepadCapture` is the one client where this is naturally per pad
already: `openSlot` computes `manager.declaredKind(for:)` and puts it in
`slot.pref`, so the question is answered where the pad is opened rather than on
every sample. `GamepadType.motionReaches(declared:asked:resolved:)` is static and
pure so it can be tested without a live session; the connection's instance method
fills in the two halves it owns, and `requestedGamepad` is stored beside
`resolvedGamepad` for the same reason it exists in the Rust client — the echo is
only this pad's answer when the pad declared what we asked for.
Where Apple differs from the SDL client, and better: it never powers the IMU. The
existing code already declined to activate sensors when forwarding was off,
reasoning that with nothing to forward there is no reason to make the pad stream
gyro over Bluetooth and burn its battery — `closeSlot` is careful to power them
back down for exactly that reason. A host that built this pad a backend without a
motion plane is the same situation, so it takes the same branch. No per-sample
check, no handler attached, and a DualSense in an X-Box-class session stops paying
for a sensor nobody reads.
The hint fires only for a pad that really has a gyro (`motion.hasRotationRate`).
A gravity-only GCMotion — what an X-Box controller exposes — would otherwise
produce a notice about a feature the player never had. That is a narrower
condition than the capture path itself uses, deliberately: making the capture
gate agree is G13's job and its own change.
The badge sits in the bottom-centre stack with the muted-mic badge and the
start-of-stream banner, at every stats tier and with the overlay off, because
this is not a statistic. Unlike the mic badge it is not a control: the setting is
not reachable mid-stream on every platform and applies from the next session
anyway. So it states the fact, names the setting, and leaves after the banner's
same 6 s. Every platform including tvOS — a DualSense on an Apple TV is an
ordinary way to play, and is exactly the pad this happens to. The model owns the
expiry rather than the view, so a second pad's hint replaces the first cleanly
instead of stacking, and ending the session cancels a pending clear rather than
carrying a stale hint into the next stream.
Non-vacuity proven by mutation, not assumed: collapsing the predicate to
`resolved.hasMotion` fails 4 assertions, including the mixed-pad row that is the
whole reason it is not a session-level check. The table mirrors the Rust one row
for row — a client that disagrees with the host here either kills a working gyro
or streams ~250 Hz into a void, and both are silent.
Gate: macOS `swift build` + the FULL suite (210 tests, 5 skipped, 0 failures) with
the two new cases observed in the run's own output, and the iOS-triple typecheck
green (`arm64-apple-ios17.0`, iOS slices + hand-assembled xcframework per the
memory recipe) — the badge and the overlay it joins are on every platform, so the
macOS build alone would not have covered them. tvOS remains unverifiable from
this Mac; the badge deliberately reuses the neighbouring banner's shape rather
than introducing anything tvOS-specific.
The revert un-reverts, on measurement: with the per-direction stamp sets
(render = the pad-proven PCM16-device/float-mix stereo split, capture =
device-format only), micpitch reads 440 Hz in as 440 Hz out at exact
peak. The octave-low voice was the driver DEFAULT endpoints disagreeing
(stereo render vs mono capture), never a raw-crossing design. The user
called the wrong verdict — the pad program 4ch success was the
counter-evidence that reopened the case.
Every 4K AV1 frame this host encoded reached the wire truncated to its
first tile, and had since AV1 was wired up. Measured on .21 (RTX 5070 Ti,
4K60, split AUTO): each access unit carried a frame header declaring two
tile rows and a single Tile Group OBU with tg_start = tg_end = 0, so
libdav1d rejected 835 of 836 AUs with "Error parsing frame header".
NVIDIA's hardware decoder accepts the truncated stream, which is why
native Vulkan Video looked healthy at 60 fps while both conformant
software decoders — rav1d in-tree and libdav1d out-of-tree — refused
every frame and clients fell to a black screen.
The two halves of sub-frame readback are armed by different conditions.
build_init_params arms the WRITER (enableSubFrameWrite +
reportSliceOffsets) from subframe_on alone; the chunked READER
additionally requires slices >= 2, and resolve_slices returns 1 for AV1
unconditionally — before the PUNKTFUNK_NVENC_SLICES override is even
read, because AV1 partitions via tiles rather than slices. So an AV1
session asked the driver to publish its output tile by tile and then took
only the first tile with one blocking lock_bitstream.
resolve_split_subframe — the one arbitration point both direct-SDK
backends already call — now disarms sub-frame for AV1 and returns
split_mode untouched, so AV1 keeps every engine split encode gives it.
Arming the reader instead is not a drop-in alternative: poll_chunk cuts
at bitstreamSizeInBytes on the reasoning that "slices are contiguous
Annex-B", which AV1's OBUs are not.
With sub-frame disarmed and split still AUTO, the same session decodes
654/654 frames clean through libdav1d.
The test that pinned this as correct (av1_untouched, "both features are
legal together") is replaced by one that pins the disarm, and by one that
checks the reader's gate against the writer's — the comparison nothing
made. The Linux latch comment claiming the two "can't disagree" is
corrected; that claim is what made this invisible.
Live bisect on a fresh endpoint: the mix/host format keys are
RENDER-engine properties — stamped onto a capture endpoint they broke its
shared-mode graph (IsFormatSupported reported 2ch/48k OK while Initialize
failed 0x88890008 on a once-stamped fresh endpoint; unstamped it opened
fine, S3). The capture now gets ONLY the device-format key — the knob
mmsys.cpl itself writes — declaring the stereo the pins actually accept.
Supersedes the check 77797a9e shipped an hour ago. The suppression, the
log-once, and the "unknown must not suppress" rule all stand; the field it reads
does not.
77797a9e read `Welcome.gamepad` — the backend the host resolved for the SESSION
— and stopped sending motion when it had no motion plane. But the host does not
build pads from that. It builds each virtual device from that pad's own
`GamepadArrival` (`Pads::set_kind`) and falls back to the session default only
for a pad that never declares one, which is precisely why `declared_kind` exists
and why its doc comment says an explicit setting has to be re-declared per pad.
So the check had a false negative, and it is an ordinary living-room setup. Under
"Automatic" the Hello carries the ACTIVE pad's kind (`auto_pref`), so a couch
with an X-Box pad on slot 0 and a DualSense on slot 1 echoes Xbox360 — while the
host, reading pad 1's arrival, builds it a DualSense with a working motion plane.
The old check read the echo, saw no motion plane, and killed pad 1's gyro. That
is the exact failure 77797a9e's own commit message names as the worse of the two
("a false negative kills working motion"), introduced by the fix for the other
one.
The question is per pad, so the slot now carries what it declared, beside the
physical `pref` it already held. The two are deliberately separate fields
answering different questions: `pref` is the controller in the user's hands, which
is what the local feedback paths must keep reading, and `declared` is the one the
host is pretending to have.
Three facts decide the predicate, and they are written out in
`pad_motion_reaches` rather than at the call site because all three clients need
the same reasoning:
- the echo is not this pad's answer when the pad declared something else;
- the host FOLDS what it cannot build — a Switch Pro on Windows, any UHID backend
on a host whose /dev/uhid is unusable — and nothing client-side can predict it;
- but the echo IS one observed sample of that fold, for the kind the Hello asked
about, so it is authoritative for a pad that declared exactly that.
Hence: trust the echo when declared == asked, else fall back to the declaration.
That keeps both motivating cases — a generic pad under Automatic (declares X-Box
360, suppressed, the sweep's H5c) and an explicit Switch Pro folded to X-Box 360
by a Windows host (declared == asked, so the echo catches it, H5d) — where either
field alone gets one of them wrong. `requested_gamepad` is kept on the client
next to `resolved_gamepad` for this: the pair is what makes the echo usable per
pad, and a lone field would only tempt the next reader back into the session-level
question.
The residual gap is a pad whose declared kind differs from the session's AND gets
folded: we keep sending and the host keeps dropping. That is the direction to be
wrong in, and it is what the session-level check was worth in the first place —
wasted datagrams, not a dead gyro.
Non-vacuity proven both directions rather than assumed. Reverting to
`resolved.has_motion()` fails on the mixed-pad row; reverting to
`declared.has_motion()` (no echo at all) fails on the Switch-Pro-on-Windows row.
Each case in the table is a session someone can actually sit down to, and the
comment on each says which of the three inputs decides it.
Gate (Linux CI image, pf-lxcheck2): fmt, `build -p punktfunk-core`, `build -p
pf-client-core`, `clippy --locked --all-targets -D warnings`, and both test
suites — green, with the new case observed in the run's own `... ok` line rather
than inferred from a green gate, and pf-client-core's 163 unchanged.
Measured resolution of the 0x88890008 mystery: IsFormatSupported said the
capture accepts 2ch/48k shared while Initialize kept failing — because
the probe itself had switched to a MONO ask for frequency counting, and
this stack does not bridge channel counts on capture even under
autoconvert. Every unopenable-endpoint verdict after that switch was the
instrument, not the endpoint. Stereo ask restored; crossings counted on
channel 0.
Exclusive+shared IsFormatSupported across {1,2}ch x {16,32}bit x
{44.1,48,96}kHz on both minted mic pins. Interrogates the DRIVER,
bypassing every endpoint-store stamping question: what the pins truly
accept decides whether the mic leg has any coherent configuration, and
whether an exclusive-mode mono open is an escape hatch. (The pad program
made its own breakthrough with exactly this instrument on the sibling
SSS driver.)
The user challenged the format-locked-pins verdict, and the pad program
is the counter-evidence: it hit the SAME 0x88890008 unopenable-endpoint
signature and cured it with a COHERENT stamp set, after which the same
driver family served 4ch happily. This branch previous attempts were
contaminated twice over — a float device-format (the pad bisect proved
the split must be PCM16 device / float mix+host) and no
AudioEndpointBuilder restart (Restart-Service Audiosrv never touches its
dependency, so endpoint configs were never rebuilt). Both mic endpoints
now get one identical coherent stereo set; the octave-low hypothesis
shifts from "raw crossing by design" to "the two endpoint stores
disagreed (stereo render default vs mono capture default)".
ROOT CAUSE, from the reporter's device log:
16:25:49.093 mic capture: 48000 Hz, 1 ch <- tap installed, format fine
16:25:49.235 audio engines joined - voice processing active
... 13 s of session, no errors, and the 10 s silence verdict NEVER fires
The engine started clean and the tap was installed against a valid
format - so neither the format timing nor the encoder was the fault. The
tripwire fires after ten seconds of CAPTURED frames and never fired
across a 13-second session: the tap received nothing at all.
Because the capture side must be pulled, and only the render graph pulls
anything. On the combined engine the input node carried a tap and no
connection, so it was not in the graph and nobody drove it: the IO unit
came up (the recording indicator lit for a beat, then went out as the
input went idle) and not one buffer ever reached the tap. No error, no
failed start - a session that quietly sent no microphone.
The input now runs through a silent sink into the main mixer, which is
what Apple's own voice-processing sample does. outputVolume = 0 because
the mic must reach the graph and never the speaker. The split path never
needed this - a capture-only engine has the input node AS its graph - so
this broke exactly when the combined topology became the default.
Verified: swift build (macOS), swift build --triple arm64-apple-ios17.0,
swift test 208 passed. Awaiting the reporter's on-device confirmation.
Supersedes the parse gate in 26b0819f. The off-thread read, the claim token, the
teardown ordering and its bounded wait all stand — only what happens in the gap
changes.
26b0819f held every report back until the calibration read came home, so a pad
that stalled on EP0 could feel dead for up to the link's 250 ms timeout: no
buttons, no sticks, nothing. Reports are now forwarded immediately and their
motion scaled by the nominal calibration until the real one lands.
That gap is exactly the behaviour that shipped before f6de620f — acceleration
~18% short, gyro unscaled — for about a millisecond. Nobody can feel that. A
controller that ignores a button press for a quarter of a second is not in the
same category, and it is the only one of the two a user would ever report.
It is also the safer of the two conservatisms available here. The rejected third
option, forwarding motion as zeroes until the real numbers arrive, would have the
host read a still pad as being in free fall — a lie about the physical world
rather than an imprecision about it. The nominal constants are merely a slightly
wrong scale.
The token is more load-bearing under this, not less. With a gate, an unpublished
calibration meant "parse nothing"; now it means "scale nominally", so begin()
clearing the previous pad's value is the whole reason a re-claim falls back to
the nominal constants instead of silently inheriting factory numbers belonging to
a different unit — which are, in general, further off than nominal. The fallback
therefore lives in the hand-off itself (MotionCalHandoff.effective) rather than
as an elvis at the call site: restoring the gate now means changing the type's
API, not deleting three characters in onReport.
The tests moved with the contract. They assert the nominal calibration is what is
in effect during the gap, rather than merely that the slot is empty — an empty
slot is now compatible with either behaviour, so asserting on it would have let
a regression pass. Added the case the change exists for: the same raw report,
parsed either side of publication, forwards identical buttons and sticks while
its gyro and acceleration convert differently. Mutation-checked three ways —
dropping the nominal fallback fails all five cases, dropping begin's clear fails
the inheritance case, dropping the token check fails three.
Gate: `:kit:compileDebugKotlin`, `:kit:testDebugUnitTest` and
`:app:compileDebugKotlin` green on a forced clean rerun, 62 cases across the
module, 0 failed, with the five hand-off cases read back out of the JUnit XML.
The on-glass re-verification f6de620f owes is still owed and unchanged.
Supersedes the synchronous calibration read f6de620f shipped an hour ago. The
ordering it protected is kept; the blocking it cost is not.
f6de620f read the pad's calibration inline in DsCapture.startUsb, which runs on
the main thread — the stream's setup path, and the USB-permission broadcast. The
read is a blocking EP0 control transfer: a pad that is there answers in about a
millisecond, but a pad that is stalling takes the link's whole 250 ms write
timeout, and either way the interface was waiting on a controller. That is the
wrong thread for it.
It now runs on its own daemon thread, one per claim, named pf-ds-cal — the same
shape HidUsbLink already uses for its reader rather than a second style. A
pathological stall now delays the pad's motion by a moment instead of freezing
the UI.
What kept the ordering honest before was "assign the calibration before `model`",
since `model` is what lets the link thread into the parse. That reasoning stands,
so the gate simply moved: MotionCalHandoff holds the claim's calibration, starts
null, and onReport parses nothing until it lands. No report is ever scaled by the
last pad's numbers — those are per unit — nor by the nominal fallback the real
read is about to replace. Dropping the first millisecond of a capture costs
nothing: the reports carry absolute state, so the next one says everything the
dropped one would have.
The calibration is what got deferred, not `model`, and that is deliberate.
Keeping `model` synchronous keeps isActive, the teardown writes, the feedback
sinks and the active-changed true/false pairing meaning exactly what they meant
yesterday — and, more to the point, it makes a late completion structurally
unable to resurrect a dead capture. A straggler can only ever publish a
calibration, and nothing is parsed while `model` is null.
Teardown, which is where this sort of change actually bites. Both stop() and the
unplug path end the claim before they close anything: ending burns the token, so
a read that lands afterwards publishes nothing and says so in the log. They then
wait, bounded at 500 ms and normally already over, for the read to let go of the
connection they are about to close — closing a descriptor with a transfer in
flight pulls it out from under the kernel, the same rule the pad-audio borrow
follows. It cannot deadlock: the reading thread blocks on the EP0 transfer and on
the hand-off's own monitor, never on anything a teardown holds. If a pad has
stopped answering entirely the wait elapses and teardown proceeds regardless,
which is the same exposure the feedback writes already carry and better than an
interface that never comes back.
Tested where it is testable. MotionCalHandoff is the piece that carries the
hazard and it is pure, so it has its own test: nothing is visible until the read
lands, a read that outlived its claim publishes nothing, a re-claim never
inherits the previous pad's calibration, and a doubled end still refuses every
outstanding token. Mutation-checked both ways — deleting the token check fails 3
of them, deleting begin's clear fails the fourth.
Not covered: DsCapture's own claim/teardown ordering is not unit-testable in this
module — there is no Robolectric, and the class builds a main-Looper Handler and
needs a UsbManager — so it is argued in comments rather than pinned. The on-glass
re-verification f6de620f owes is unchanged and still owed.
Gate: `:kit:compileDebugKotlin`, `:kit:testDebugUnitTest` (61 cases across the
module, 0 failed) and `:app:compileDebugKotlin` green, with the four new cases
confirmed present in the JUnit XML rather than assumed from a green build.
Field report: mic uplink dead on iOS, iPadOS and macOS alike, while
Android on the same host works - so the host and the wire are fine.
Two defects in the combined (voice-processing) engine, which became the
default on all three Apple platforms a week ago and has never run on a
device - CI only runs swift test on macOS, and the loopback test counts
datagrams without decoding them.
- The tap read the input format before the engine was prepared. Enabling
voice processing swaps the engine's IO unit for the VPIO one and
renegotiates its formats; until prepare() the input node can still
report the pre-swap state, 0 Hz / 0 channels included, which
installMicTap correctly refuses as 'no usable input device'. Both
topologies now prepare first, so the chain is built against what the
voice processor actually emits.
- A mic chain that failed on the voice-processed engine took the whole
uplink down for the session: that arm fell back to playback ONLY. The
sibling failure a few lines above - the voice processor refusing to
engage at all - already falls back to the split path, which is a
working mic without echo cancellation. Both arms do that now. The mic
outranks the AEC.
Not reproduced locally (no Punktfunk entries in this Mac's log store,
and collecting the device's log needs root), so this is a strong
inference plus one proven logic defect rather than a confirmed fix. If
it persists, Console filtered to subsystem io.unom.punktfunk / category
audio names the stage: 'mic capture: N Hz' then 'audio engines joined'
then, 10 s in, either 'mic uplink OK - peak ...' or the SILENCE warning.
Follow-up worth doing separately: nothing reports whether the uplink
actually opened, so the HUD offers a Mute Microphone button over a
session sending nothing. Android gates that on a real micRunning signal.
Final pitch-probe verdict on the SSM driver pair: the render pin is
stereo-only, the capture pin mono-only (stamping either differently makes
the endpoint unopenable), and the crossing between them is a RAW byte
pass — so voice fed through the render endpoint reads back an octave low
and no format stamp can fix it. S3 peak-based PASS = false pass; per the
design doc revert clause the mic falls back to the name ladder (a virtual
cable), pending the user re-decision. The SPEAKERS substrate keeps tier-0
(no driver crossing — a plain engine loopback tap, measured clean).
minted_ids() publishes speakers only; the mic endpoints stay minted and
recorded (provisioned()) for the micpitch probe and a possible future
non-render transport, and their format stamps now pin each side to its
pin one true format — healing the endpoints this branch earlier
mis-stamped.
Second measurement round: the driver render pin is STEREO-ONLY — the
mono render stamp turned the endpoint unopenable (0x88890008 on every
open, the incoherent-stamp signature the pad program documented). Since
the crossing is raw, the coherent choice inverts: the CAPTURE side now
declares the stereo float stream that actually crosses (fixing the
octave-low voice), and the render has its stereo float default stamped
explicitly — pinning the pair AND healing any endpoint a previous build
left mono-stamped.
Measured with the new pitch probe: 440 Hz into the minted mic render came
back as 220 Hz off its capture side. The driver forwards the render
stream RAW into its mono capture, so a stereo-declared render (the
driver-default we inherited) turns every stereo frame into two mono
samples — half speed, octave down, exactly the field report. The mic
render now gets a coherent MONO 48 kHz format set stamped alongside its
name (PCM16 device format + float mix/host formats), making the engine
downmix before the driver crossing. The mic pump keeps pushing stereo;
shared-mode autoconvert handles the rest.
Field report through the minted microphone: voice plays back an octave
low. Peaks are pitch-blind — S3 passed while a potential half-rate link
hid in the numbers (288k samples fits both the honest and the half-speed
story). Every probe measurement now estimates the dominant frequency by
zero crossings over the signal span, and `audio-probe micpitch` runs the
decisive experiment against the LIVE minted pair: 440 Hz in, frequency
out — ~440 = pair innocent, ~220 = the stereo render stream is forwarded
raw into the mono capture.
Making the HUD concentric with the physical display corner had no upper
bound, so a modern phone (~62 pt of display radius) asked for a 48 pt
corner on a card whose lines sit 10 pt from the edge. A corner of radius
r pulls the edge inward by r - sqrt(r^2 - (r-y)^2) at distance y below
the top: at the first line that is ~19 pt, so the top and bottom lines
rendered INSIDE the arc.
Concentricity is only a virtue while the radius is small next to the
card. The radius is now capped at 28 (devices asking for less still get
a truly concentric corner) and the iOS content padding scales with it at
0.45*r, which leaves ~4.6 pt of arc against 12.6 pt of padding at the
cap. The card grows by under 3 pt a side; the compact pill is unchanged.
G14/G16 leg 3. This supersedes the nominal constant 0e40b374 shipped, which was
always labelled a stopgap.
Measured on glass 2026-08-07: a DualSense over USB into an Android phone,
streaming to a Linux host, flat and face up, arrived as |accel| = 0.811 g where
1.000 was owed. The parse forwarded the pad's raw i16s verbatim, and raw device
units are not wire units. 0e40b374 rescaled acceleration by the nominal
10000/8192 and deliberately left gyro alone, because a constant provably cannot
fix gyro: the same still average showed this unit's accel calibration is
near-identity (~1% off) while its gyro's emphatically is not — a near-identity
gyro calibration would imply 1024 LSB per deg/s, i.e. ±32 deg/s full scale, which
no controller has. That scale is per unit, and the only thing that knows it is
the pad.
So the client now asks. HidUsbLink grows a GET_REPORT path — EP0, the exact
mirror of the SET_REPORT it already had — and DsCapture reads the pad's IMU
calibration feature report ONCE, while claiming it: 0x05 / 41 B on a DualSense or
Edge, 0x02 / 37 B on a USB DualShock 4. DsDevice.MotionCal then applies
hid-playstation's own arithmetic per axis, which is the same math the host's
contract test (crates/pf-inject/tests/motion_contract.rs, SonyImuCalibration)
reads from the other end: gyro raw × speed_2x × 20 / (|plus−bias| + |minus−bias|),
accel (raw − (plus − range/2)) × 20000 / range. Long arithmetic, because the gyro
multiplier overflows an Int, and clamped, because both are >1 multipliers and a
full-scale flick would otherwise wrap the i16 into a motion in the opposite
direction. Reading the blob also removes acceleration's residual ~1% factory bias
that the nominal constant left behind.
Once at claim and never per report. EP0 is independent of the interrupt endpoints
so the read is safe alongside the reader thread, but a blocking control transfer
in the report path would wreck capture latency, and the calibration is fixed for
the life of the connection anyway. The capture logs the derived resolutions, which
is the discriminator for whether a blob was read at all: a real pad declares ≈16
LSB per deg/s, the fallback reads back as exactly 20.
A pad that refuses, answers short, or declares zeroes (a clone, a broken unit)
keeps today's behaviour per axis — nominal accel, gyro straight through. Nothing
here ever zeroes motion: slightly mis-scaled beats silent.
Not covered. The axis frame is still untouched: this leg puts gravity on Y where
the Apple leg put it on Z, so at least one client's frame is wrong, and settling it
needs the bare-metal Linux reference reading G16 step 1 calls for. Rescaling is
frame-independent, so it stands however that resolves — remapping is not, so it
stays out. Bluetooth's grouped plus/minus layout is not implemented either: this
path is USB-only by construction (Android exposes no raw path to a Classic pad),
and a half-used generalisation would be a latent bug rather than a feature.
Gate: `:kit:compileDebugKotlin` + `:kit:testDebugUnitTest` green, 16 DsDeviceTest
cases run 0 failed, and the five new ones were confirmed present in the JUnit XML
rather than merely compiled. Non-vacuity checked by mutation — perturbing the gyro
conversion fails 6 tests, including all four new ones that assert a number.
On-glass re-verification owed, on the rig that measured the defect (DualSense →
USB → phone → 192.168.1.21): at rest |a| = 1.00 g exactly via ~/gyroscope.py, and
a nominal 90 deg yaw integrating to ~90 deg via ~/integrate.py — the same 90 deg
that read ~62.7 deg before this change.
Three things the Intel Arc measurement showed were wrong or unhelpful in the refusal
path.
The message named NV12 whatever the stream was. A Main 10 session refused over P010
was told about NV12, which sends the reader to look up the wrong format's support.
Both variants now carry the format the driver's own entry reported.
A missing SAMPLED now says what it costs. "does not advertise usage SAMPLED" is
accurate and tells a field reporter nothing: the consequence is that no shader can
read this device's decoded pictures, so the zero-copy path cannot exist on it at all
— which is a different conversation from a device that is merely slower. The line
points at --probe-decode for the driver's own words.
And the probe's second opinion no longer claims to be one. Measured on both vendors,
vkGetPhysicalDeviceImageFormatProperties2 answers "creatable" for combinations the
video-format query rejects — on NVIDIA too, for SAMPLED alone, which is not a legal
video image usage at all. So it does not honour the chained profile list and must not
be read as permission; it is still printed, because otherwise everyone who reads a
refusal asks the question again, but it is labelled as not authority.
Also names the three video ENCODE usage bits, which NVIDIA advertises on decode
pictures and the probe was printing as "unrecognised 0xC000".
Structural, because tuning the transform values was treating a symptom.
A scroll transition derives its phase from the geometry of the view it
wraps, and the entrance was wrapping each card on the OUTSIDE - so it
moved the very thing the transition measures. Every card read as far
from centre for the whole travel, phase pinned at fully receded, and the
centred card only collapsed into its focused look as the entrance ended.
That collapse was the jump; shrinking the offset last round only made it
smaller.
The card builder now hands each caller its own CardEntrance and both the
launcher and the coverflow apply it BENEATH their .scrollTransition. The
transition measures a card that never moves and composes its scale and
rotation on top of the entrance's, so the two can no longer fight - and
the fuller travel is back (34 pt rise) now that the geometry constraint
that forced it down to 16 is gone.
The focused card jumping into its correct state at the end of the
entrance was the entrance's own geometry. The caller's .scrollTransition
reads the geometry of the view underneath the entrance's transforms, so
a card shoved 58 pt down and hinged on its leading edge spent the whole
travel reported as far from centre - phase pinned at fully receded - and
only collapsed to identity as the card came home. That collapse IS the
jump, and it explains why it looked timing-dependent rather than simply
broken.
Now the rotation is about the card's centre (it turns in place instead
of swinging sideways out of position) and the rise is 16 pt, inside the
strip's own vertical slack, so nothing the entrance does moves a card
away from where the scroll view thinks it is. The entrance also waits a
couple of frames for real layout - the GeometryReader's first pass can
report no width, so there is nothing to centre on yet - and the
transaction override from the previous round is gone: it was not the
cause, and nil-ing inherited animation could have made navigation
snappier than intended.
The Intel Arc refusal moved one step down the caps query and stopped again: the
coincide NV12 entry does not advertise SAMPLED. That sentence is punktfunk's, not
the driver's, and the last two times a conclusion was drawn from a sentence of ours
the conclusion was wrong.
So --probe-decode now prints the driver's own answers instead. For every profile the
client can negotiate (H.264 High, H.265 Main and Main 10, AV1 Main 8- and 10-bit) it
asks vkGetPhysicalDeviceVideoFormatPropertiesKHR in six usage combinations — the
three the image pools really create with, plus DPB|DST without sampling, SAMPLED
alone and DST alone, which are what localise a refusal to a half. Each answer is
printed as the driver gave it: format, usage and create flags named AND in hex with
unrecognised bits called out, image type, tiling. A failed query prints its VkResult
rather than vanishing into an empty list.
It goes through pf-vkdecode's own query rather than a copy of it, which meant
splitting query_formats into a physical-device form — the call never needed the
VkDevice the old signature demanded. VideoFormat gains imageType and imageTiling to
carry the whole record; VUID-VkImageCreateInfo-pNext-06811 compares both for
equality, so they were being assumed rather than read.
And because a driver that under-reports usage would be indistinguishable from one
that genuinely lacks it, the probe asks a second, independent question —
vkGetPhysicalDeviceImageFormatProperties2 over the same profile list — and prints it
only where the two disagree. A disagreement is the finding.
No behaviour change to any decode path: derivation reads the same fields it did.
Field-measured necessity, not cosmetics: unstamped, the minted instances
read 'Lautsprecher (2- Steam Streaming Microphone)' and even the box's
owner picked the wrong device out of the Sound settings zoo (as did the
S1 probe's name match before it). The provider now stamps device-desc +
device-name through the pad program's proven machinery — write_stamps/
stamps_served, extracted from the pad-only stamp functions — with the
same store-first/registry-fallback routes and settle/re-pass discipline.
Names only: a wider stamp set makes AudioEndpointBuilder re-mint the
endpoint under a new GUID (measured on pads). Stamping is best-effort
(SYSTEM ACL route); the wiring never depends on names — identity stays
the recorded id.
The strip entrance is one animated progress value now, not a Bool behind
per-card .animation modifiers. Those modifiers wrap the caller's card -
INCLUDING its .scrollTransition - so a delayed spring flipping while the
scroll view was still settling captured the transition's own per-frame
phase updates and stranded the centred card half-receded until the next
scroll re-drove it. That was the 'only navigating fixes it' report, and
the race with load speed was the same thing.
CardEntrance is now a ViewModifier + Animatable: it slices its own
window out of one master clock the carousel animates 0 -> 1, so every
transform is a pure function of an interpolated Double and no animation
modifier wraps a card at all. Benign failure mode too - progress
reaching 1 without animating leaves each card at exact identity rather
than stranded. The entrance also moved inside .frame(width:) so a scroll
target's geometry never depends on what its card is doing, and the
non-tvOS branch states its .id explicitly.
One leak remained after that: withAnimation sets its animation on the
whole TRANSACTION, so the scroll view's initial centring still inherited
the 1 s linear clock and the focused card only reached its correct look
as that clock ran out - arriving as a jump. The card subtree now clears
the inherited animation, so its phase lands per frame while the
entrance's own transforms (driven by animatableData, not by the
transaction) keep running.
G16 leg 2. A DualSense over USB to an Android phone, streaming to a Linux host,
flat and face up: |accel| = 0.811 g where 1.000 is owed. Magnitude is
frame-invariant, so this is unambiguous regardless of the separate axis question
below, and it came from a 27-second static average — no sampling error in it.
`DsDevice` said so plainly: "Gyro/accel stay in raw device units". It read the
i16s out of the pad's report and forwarded them verbatim. But raw device units
are not wire units — the wire is fixed at 10000 LSB/g and the pads' native
resolution is the 8192 that hid-playstation calls DS_ACC_RES_PER_G. 8192/10000 =
0.819 predicted against 0.811 measured. Acceleration is now rescaled on both the
DualSense and DualShock 4 parse paths, clamped because the multiplier is >1 and
a real near-full-scale slam would otherwise wrap the i16 into an impossible
acceleration in the opposite direction.
Two things deliberately NOT done.
Gyro is left alone. It is almost certainly low by the same mechanism, but it
cannot be corrected with a nominal constant the way acceleration can: the still
average shows this pad's accel calibration is near-identity (~1% off), while the
gyro's emphatically is not — a near-identity gyro calibration would imply
1024 LSB per deg/s, i.e. ±32 deg/s full scale, which no controller has. Fixing
gyro means reading the pad's calibration feature report and applying its own
numbers, which also removes acceleration's residual 1% bias. `HidUsbLink` can
SET_REPORT but has no GET_REPORT path yet, so that is a real change rather than
a constant, and it is owed.
I tried to pin the gyro factor by integrating the on-glass rotations instead: a
nominal 90 deg yaw integrated to ~88.5 deg through the Apple client (correct)
and ~62.7 deg through Android. Directionally consistent, but the readout samples
at 5 Hz and a ~1 s rotation is badly undersampled, so that ratio is not a
constant anyone should ship. Recorded, not used.
The axis frame is also left alone. This leg puts gravity on Y where the Apple
leg put it on Z, so at least one client's frame is wrong — but Android forwards
the pad's own axis order un-remapped, which makes its reading evidence about the
hardware rather than about us, and resolving it needs the bare-metal reference
reading G16 step 1 calls for. Every bare-metal Linux box was unreachable
(Deck down, HTPC down, .25 is another KVM guest). Rescaling does not touch axis
order, so this fix stands however that resolves.
Gate: `:kit:compileDebugKotlin` and `:kit:testDebugUnitTest` green, JNI libs
built clean at the API-28 floor across 3 ABIs. On-glass re-verification owed:
re-run the at-rest reading and expect 0.99-1.00 g.
Three defects behind an entrance that read as a card sliding up:
- The centred card never rotated. The stagger fans out from an anchor,
and the anchor was given side 0 = no rotation - but the anchor IS the
card the eye is on, so the single most visible card only rose. Side is
never 0 now; every card turns.
- The swing happened while the card was invisible. Opacity shared the
transform's spring, so the card spent its whole rotation at near-zero
alpha and only the last few degrees showed. The fade now runs on its
own 0.22 s curve (a second .animation governs only the modifiers above
it) while the transform springs over ~0.6 s. The travel is deeper too
- 0.74 scale, 64 degrees, 58 pt - and the rotation sign now matches the
coverflow's own recede, so a card unwinds INTO its resting angle
instead of swinging against it.
- It fired before the art existed. Cards swung in as grey placeholders
and filled with artwork afterwards. PosterImage reports when a cover
settles (art loaded, or candidates exhausted), the coverflow counts
the first few, and GamepadCarousel holds its entrance on a
contentReady gate - with a 700 ms backstop so a slow or artless
library still animates.
The strip entrance never ran in the library, for two reasons:
- The trigger was lost. Flipping the state inside onAppear puts the
change in the SAME transaction as the view's insertion, where SwiftUI
runs with animations disabled. The launcher got away with it; the
library's strip mounts late - only once the fetch lands - and lost
every time. The flip now defers one runloop turn, so it is an ordinary
animated state change.
- The art snapped in behind it. Covers hard-swapped from grey
placeholder to image, so even a working entrance was followed by a run
of cards popping to artwork after the strip had settled. PosterImage
cross-fades now (the touch grid inherits it).
And the entrance is 3D: a card starts turned away on the drum, small,
low and invisible, then swings flat, grows and rises on an overshooting
spring. Cards left of the anchor hinge on their trailing edge and cards
right of it on their leading one, so the strip FANS OPEN from the cursor
instead of sweeping past it - the same hinge-and-perspective language
the coverflow's own recede speaks, so arriving and scrolling read as one
object. Reduce Motion still drops every bit of travel.
Intel Arc never used Vulkan Video decode on Windows. The rung refused every
session with "driver advertises neither DPB_AND_OUTPUT_COINCIDE nor DISTINCT"
and fell back to D3D11VA — and that refusal was ours.
vkGetPhysicalDeviceVideoCapabilitiesKHR was called with the codec capability
struct chained BEFORE VkVideoDecodeCapabilitiesKHR (push_next prepends, so the
chain was caps -> h265_caps -> decode_caps). On Arc/Windows 101.8724 the driver
fills those two by POSITION, not by sType, and returned them SWAPPED. Measured,
on glass, both ways:
before: decode_flags_raw=12 max_level_idc=1
after: decode_flags_raw=1 max_level_idc=12
12 is STD_VIDEO_H265_LEVEL_IDC_6_2 and 1 is DPB_AND_OUTPUT_COINCIDE. We were
reading an H.265 level as a decode-capability bitmask; 12 contains neither 0x1
nor 0x2, so the check concluded the device had no DPB mode. It had one all along.
The base struct was fully populated throughout — 15 DPB slots, 8192x8192 max
extent — which is what gave the lie away: a driver that answers in that much
detail is not declining.
NVIDIA and RADV dispatch by sType and do not care about the order, which is
exactly why the fleet stayed green and this reached the field. Both orders are
spec-legal for us to write; only one survives a driver that assumes the
conventional one, and the conventional one — decode caps first, as every Vulkan
sample writes it — is now what all three codecs use.
⚠ This does NOT yet give the Arc Vulkan Video. It moves the refusal one step
down the same function: the device advertises only COINCIDE (no DISTINCT), and
its NV12 coincide entry does not advertise SAMPLED usage, which the zero-copy
presenter path needs. Whether that is a second bug of ours or a real Intel
constraint is not yet established, and this commit does not claim it either way.
Found because the user disbelieved my "Intel driver bug" conclusion. He was
right: I had reasoned from our own error message, which is the same circularity
the caps logging added in fb1a0a61/a183cac8 now exists to break.
Gates: fmt clean; clippy -D warnings; 187 pf-vkdecode tests. The GPU parity legs
that cover this code cannot run here (no GPU on the build host) — the evidence
is the on-glass A/B above.
Two more from the on-glass pass:
- The coverflow's store/source chip only showed its background on the
centred cover. Same mechanism as the tray blur: a card rides a
scrollTransition that composites it with opacity < 1 and a 3D
rotation, and a material cannot sample a backdrop through an offscreen
composite - so the frost stayed blank everywhere except the one card
sitting at exactly full opacity. The coverflow's chip is a flat wash
now (StoreBadge gains `solid`), which has no backdrop to sample and
is therefore simply always there. The touch grid keeps its material -
its cards carry no transform, so its frost samples fine.
- Host cards and library covers now arrive with the strip instead of
being there: each card rises out of a fade on a lightly overshooting
spring, delayed by its distance from the cursor, so the strip
assembles outward from where the eye already is. Implemented once in
GamepadCarousel, so the launcher and the coverflow inherit it
together. Transforms only - snapping, the callers' own
scrollTransition and the tvOS focus engine are untouched - and Reduce
Motion drops the travel for a plain unstaggered cross-fade.
G16, first result. A DualSense paired to an iPhone, streaming to a Linux host,
lying flat and face up: hid-playstation decoded z = −0.99 g where a DualSense
owes +1.00. Vector magnitude was 1.006 g, so the scale was already correct —
this is purely direction, and it was wrong for every accelerometer sample the
Apple client has ever sent.
The cause is a convention mismatch, not a sign typo. Apple reports acceleration
as the gravity VECTOR, which points down: a device face-up on a table reads
z = −1. An accelerometer physically measures proper acceleration, and at rest
that is the +1 g normal force pushing UP — which is what a DualSense's report,
and therefore our wire, carries. The two are exact negatives. Both branches were
affected, because `m.acceleration` follows the same Apple convention as the
gravity/userAcceleration split, so reading the "raw vector" was not an escape
from it.
`rotationRate` is a true angular rate and needs no flip. The same session
confirmed that independently: rotating the pad clockwise seen from above
produced a negative yaw, which is correct under the right-hand rule about an
up-pointing Z. That asymmetry — accel wrong, gyro right — is itself evidence for
this diagnosis rather than a blanket frame error, and it is why the fix is three
negations at one site instead of a remap.
The sweep predicted this ("Apple accel plausibly INVERTED — CoreMotion gravity
-1 g vs DS +1 g up at rest") but could not confirm it without hardware. It is
now measured, and the mechanism is confirmed in the code rather than inferred
from the number.
Method, for whoever repeats it: the readout is python-evdev on the host reading
the virtual pad's own motion node, dividing by the axis `resolution` the kernel
publishes, so it prints deg/s and g. That is downstream of the calibration blob
— the same layer a game reads — which is what makes a sign error visible to a
human at all.
Two things this does NOT establish. The host was a KVM guest, so the DualSense
could not be attached natively for a side-by-side reference reading; the test
stands on the DualSense convention being a fixed property of the hardware, which
is decisive for the at-rest sign but weaker for the gyro axis ORDER. And the fix
itself is unverified on glass: confirming it needs a rebuilt client on the
device, so someone should re-run the same at-rest reading and see +1.00.
Gate: `swiftc -parse` clean. A full typecheck needs the gitignored
PunktfunkCore.xcframework assembled first and has not been run.
The Arc returned decode_flags=0b1100 = 12 with a fully populated base struct (15
DPB slots, 8192x8192 max extent). Neither COINCIDE (0x1) nor DISTINCT (0x2) is
set, and 0x4|0x8 are not defined for that field at all — but 12 IS
STD_VIDEO_H265_LEVEL_IDC_6_2, and VkVideoDecodeCapabilitiesKHR and
VkVideoDecodeH265CapabilitiesKHR have identical layouts (sType, pNext, one u32).
So the suspicion is that we are reading H.265's maxLevelIdc where the decode
flags belong. Logging both settles it: if max_level_idc comes back as 1 or 2 the
two structs are crossed, and the refusal is ours rather than the driver's.
Nothing in the caps module logged anything, so when a device refused with
"advertises neither DPB_AND_OUTPUT_COINCIDE nor DISTINCT" there was no way to
separate two very different situations that present identically as a zero: the
driver filling the chain and genuinely declaring no DPB mode, versus our own
pNext chain never reaching VkVideoDecodeCapabilitiesKHR at all.
Printing the BASE VkVideoCapabilitiesKHR beside the decode flags is the
discriminator. A populated max_dpb_slots next to decode_flags: 0 means the
driver traversed the chain and answered; zeros across both mean the query never
landed and the refusal is ours, not the driver's.
Raised by the Intel Arc result on .221, where I concluded "driver bug" on the
strength of our own code's report — which is precisely the circular reasoning
this line exists to break.
Round-3 field findings: the pop-in had retreated to the X axis alone -
the one growth still left to safe-area resolution (the landscape side
insets), which settles a beat after insertion, outside any geometry
group and outside the view's own transaction. The scrim now reaches
full-bleed purely by layout: a fixed 80 pt negative-padding overhang on
the outer edge and both sides replaces ignoresSafeArea entirely, so
every axis is deterministic from the first frame (and orientation no
longer changes the gradient's proportions). The mask's strong region
moves to 0.65 to account for the overhang leading the gradient.
And the frost reads black now, not grey: an ink.shade(0.35) wash inside
the mask sinks the material's luminance lift toward the palette's shade
- black on a dark field, palette-honest on a pale one.
Two follow-ups from the second on-glass pass:
- The tray blur's pop-in survived the geometryGroup: the full-bleed
growth (negative padding + safe-area expansion) rode the push's
transaction, and safe-area resolution sits outside a geometry group.
The scrim now pins its own geometry out of any animation - the layer
fade/slide still carries it, only its SHAPE can never animate. The
bottom overshoot grows 32 -> 72 pt (the tray sits over scrolling rows
plus the detail line; the blur influence starts well above the legend
now) and the mask holds strength longer before dissolving.
- The selected tab pill is a Liquid Glass surface (accent-tinted
through consoleGlass, material fallback pre-26/tvOS) - the strip
wears the same material language as the rows below it. The
matched-geometry travel between pills is unchanged.
Six findings from the on-device pass over #91, all iOS-facing:
- The tray blurs no longer grow into place on a push: the screen layer
resolves its internal layout (safe-area trays, the scrims' full-bleed)
in a geometryGroup BEFORE the insertion animates.
- The option band is LINEAR now, not a ring. A ring showed the first
option waiting to the right of the last one - unreachable, since
left/right clamps - and on a 2-option ring the unselected item flipped
sides with every step (the 60/120 Hz row). Positions are fixed, the
ends are the ends, and A's wrap travels back across the list. Options
other than the facing one exist only while the drum is moving, so a
long label never sits under a resting neighbour as overlapping text.
- Toggles (and the pin rows) ride the band too: Off left of On,
matching the left-off/right-on step semantics.
- The close X is gone from settings, add-host and the library - a
gamepad UI exits with B. A chromeless cancel button keeps hardware
Esc and the macOS sheet working, and the library's loading/error/empty
states gain a zero-size B listener so a controller-only user is never
trapped where the coverflow (and its B) doesn't exist yet.
- The heading is a real heading: leading-aligned with the 24 pt content
inset, 24/34 pt (was 20/30), top margin 18/28 (was 10/18) - launcher,
settings, add-host and library alike. The launcher's hidden-mirror
chip trick died with the centred title that needed it.
Verified: swift build (macOS), swift build --triple arm64-apple-ios17.0,
swift test 208 passed / 0 failed.
Found on glass, and it cost a whole session to find. PUNKTFUNK_DECODER was read
untrimmed, so "native-vulkan " — ONE trailing space — matched no arm of
native_vulkan_gate, fell through to `auto`, and on an Intel box `auto` takes
d3d11va first. The operator's pin never ran and NOTHING said so. Read against a
log, that is indistinguishable from the rung being refused for a hardware
reason, which is precisely the ambiguity the rest of this module's logging was
just rewritten to remove.
The space is not exotic. A Windows .cmd produces it for free: `echo x>> file`
keeps the space before the redirect, so every line written that way carries one.
PUNKTFUNK_VK_ADAPTER already trimmed; this did not, and the inconsistency is what
made it invisible — the GPU override obeyed while the decoder override did not.
The rule now lives in one pure function, resolve_decoder_pref, called by BOTH
readers. decode_pinned_to_software had the identical untrimmed expression, and
its own doc comment says a second reading of the same two inputs is a second
place for them to drift — fixing one and not the other would have proved it
right. Whitespace-only counts as ABSENT rather than as a pin to "", because an
exported-but-empty variable means "no override" and "" is a value the gate
happens to accept.
Tested as a pure rule (no process environment), including the end-to-end leg
that matters: the trimmed pin reaches native_vulkan_gate and is admitted. Like
the create-array tests in dee97e89 its before-state is a compile error rather
than a failing assertion, because the function is new — what it guards going
forward is real, and an editor who drops the trim fails it.
Gates: fmt clean; clippy -D warnings over pf-client-core,
punktfunk-client-session and pf-presenter in the Linux container; 164
pf-client-core tests.
Working G14/G18 turned up two sweep findings that do not survive contact with
the code. Neither is implemented; one is now guarded.
The 2026-08-07 sweep read the Triton (Steam Controller 2) usbip endpoint's
`bInterval: 1` as 125 µs — an 8 kHz duplicate storm — and the plan's G14 says to
raise it to 4 "like the Deck". That reading assumes a high-speed device, where
bInterval is the 2^(n-1) × 125 µs exponent. Both Triton devices declare
`UsbSpeed::Full`, and on a full-speed device the field is a plain frame count in
milliseconds: 1 means 1 ms, which is the 1 kHz the existing comment claims.
Raising it to 4 would mean 4 ms — a 4× cut to the motion rate a passed-through
SC2 delivers, in the name of fixing a problem it doesn't have. The endpoint now
carries the reasoning so the next reader doesn't repeat it.
G18's first bullet ("bound/rate-cap the host's rich-input channel; motion is
unbounded") is stale rather than wrong — it was true of the tree the sweep read.
Current main already routes rich input, motion included, through a 1024-deep
`sync_channel` whose `offer()` helper `try_send`s and drops on full, ending the
loop only on Disconnected. That is the same bounded-queue pattern the mic plane
adopted for security-review S6. Nothing owed.
G14's remaining bullet — DS/Deck neutral accel should read 1 g on the up axis
instead of 0 g free-fall — is deliberately NOT done here. Which axis is up is
precisely what G16's on-glass session measures: `switch_proto` documents the
wire as z-up and its neutral ships +Z, but the Deck's kernel negates Z/RZ, so
guessing would leave one backend confidently disagreeing with another. A wrong
constant is worse than the current obviously-unset 0.
Gate: fmt, build, clippy --all-targets -D warnings, and the test suites — green.
Four reworks from the first palette-era on-glass review, all iOS-facing:
- Surfaces carry the palette now, not just the text on them: ConsoleGlass
washes every tier (Liquid Glass tint, pre-26 material, tvOS material)
with ink.glass — the same colour the desktop console fills its panels
with — and the close buttons move to an ink-aware consoleGlassBackground.
The pre-26 branch also gains the focus tint it had silently dropped.
Stray literals follow: ConnectOverlay text rides ink in the console
takeover, card shadows soften on pale fields, the focused keycap reads
onAccent. The online pip stays status-green on purpose.
- The header breathes: title top padding 4/10 -> 10/18 plus shared
header-spacing and title-bottom helpers mapped from the console shell's
rhythm, applied to the launcher, settings and add-host alike, with the
add-host close X re-anchored to the title row.
- Settings, Add Host and the Library present IN PLACE on iOS: one
persistent aurora whose calm is chased (the console's bg_mix), screens
as transparent layers with the console's 0.26 s ease-out-cubic push/pop,
an input drop for the transition, and the controller handed off through
isActive — no more opaque bottom-up covers, no backdrop teardown.
macOS keeps its sheets, tvOS its focus-engine covers.
- The settings select is a real band: choice rows mount GamepadOptionBand,
a spring-driven drum (Animatable body, ring-distance wrap, neighbours
gated by focus and flight) whose retargeting spring accumulates rapid
steps into one continuous spin. Reduce Motion falls back to a plain
crossfade; toggles keep the quiet 14 pt slip.
Verified: swift build (macOS), swift build --triple arm64-apple-ios17.0,
swift test 208 passed / 0 failed. On-glass QA still owed: palette sweep on
a pale palette, transition compositing over materials, drum feel on device.
G8 of the gyro program, SDL-client half.
The `Welcome` has always carried the backend the host actually RESOLVED, which
is not necessarily the one the client asked for — Auto lands on Xbox 360 for
anything not Sony/Valve/Xbox, and a Switch Pro on a Windows host folds to X360
too. No client read the field. So a player with an 8BitDo, or a Switch Pro on
Windows, got a controller whose gyro did nothing, with nothing anywhere saying
why: the client shipped ~250 Hz of Motion datagrams and the host parsed and
discarded every one.
`GamepadPref::has_motion()` answers whether a backend has a motion plane at all.
The SDL client checks it on the first gyro sample: it logs one line naming the
resolved backend and pointing at the fix (pick a DualSense-class controller
type), then stops sending. Once per slot, not per sample — this path runs at the
pad's sensor rate.
`Auto` deliberately answers true. It means "unknown" — an old host that omitted
the echo, which may well have resolved a DualSense — and suppressing motion on
unknown would silently break working gyro, a worse failure than sending
datagrams nobody reads. The predicate is an exhaustive match so a new backend
has to state its answer rather than inherit one, and a table test pins both
halves: a false negative kills working motion, a false positive keeps the void
open, and both are silent.
Owed: the plan wants this surfaced as a one-line UI hint, not just a log line.
Apple already stores `resolvedGamepad` and Android needs the plumb; neither is
done here, and both want their own gate.
Gate (Linux CI image): fmt, build, clippy --all-targets -D warnings, and the
test suites — green, with the new capability test observed running.
--probe-decode printed its DISPLAY position and called it the
PUNKTFUNK_VK_DEVICE value. It is not. pick_device resolves that variable against
the RAW vkEnumeratePhysicalDevices order (setup.rs, `devices.get(i)`) BEFORE any
ranking runs, while the probe sorts discrete-first for readability.
Those two orders disagree precisely on the hardware this flag exists to
diagnose. pick_device's own comment records why the ranking is there: "enumeration
order puts the iGPU FIRST on some hybrids (observed: Ryzen iGPU ahead of an RTX
dGPU)". So on a hybrid laptop the number the probe printed for the iGPU could
well be the number for the dGPU — a diagnostic handing out an actionable value
that selects the other GPU, which is worse than printing none.
Measured on the Arc + RTX 3500 Ada laptop, which is also where the first output
went out with the wrong claim in it: three adapters, and the same Arc iGPU
enumerated TWICE. So AdapterDecode now carries the raw enumeration index,
captured before the sort, and the printer uses it; the "default presenter"
marker stays on the first LISTED entry, because sorted-first is what pick_device
lands on when nothing overrides.
The duplicate is why the trailing hint names PUNKTFUNK_VK_ADAPTER as the safer
knob and admits its limit: two adapters sharing a marketing name cannot be told
apart by it, and a name match resolves to whichever enumerates first. The hint
also states the thing this whole output invites a reader to get wrong — that a
capable GPU in the list does not mean the decoder will use it, because Vulkan
Video decodes on the presenter's device and PUNKTFUNK_DECODER does not move the
presenter.
Gates: fmt clean; clippy -D warnings on punktfunk-client-session and pf-presenter.
First hardware run of --probe-decode, on the RTX 5070 Ti:
driver decode ops: H.264, H.265, AV1 (0xF)
Three names, four bits. 0xF is H.264|H.265|AV1|VP9 — bit 3 is
VK_VIDEO_CODEC_OPERATION_DECODE_VP9_BIT_KHR, a real decode operation this
client has no rung for, so the name table stopped short of it and the line
looked complete while silently dropping a codec the driver had advertised.
That is the exact failure this flag exists to prevent. The whole point of
--probe-decode is that a reader can trust the words to cover the number; a mask
with an unexplained bit asks them to trust it instead. VP9 is now named (marked
as having no punktfunk rung, because advertising it as decodable would be its
own lie), and any bit beyond the four we know prints as "unrecognised bits
0x…" rather than vanishing — so the next codec Khronos adds shows up as an
unknown rather than as nothing at all.
Gates: fmt clean; clippy -D warnings on punktfunk-client-session.
G6 + G15 of the gyro program.
G6 — the UMDF gamepad driver's input path. Its timer ran at 8 ms and completed
one pended READ_REPORT per tick, so a game could observe at most ~125 Hz while
clients stream motion at ~250 Hz: every other sample was overwritten in the slot
before anything read it, and the ones that survived carried up to 8 ms of extra
latency. For gyro, a dropped sample is not a dropped frame — it is rotation that
never reaches the game.
The timer now ticks at 2 ms (about a real DualShock 4's Bluetooth cadence). Only
the cheap half runs on every tick: read the input slot, complete one pended
read. The channel handshake and the health marks stay on their historical ~8 ms,
because they cost more, nothing wants them faster, and `driver_heartbeat`'s
documented "+1 per ~8 ms tick" is what the host reads as liveness.
The same slot is a single unqueued buffer that both sides touch without a lock,
so a driver read landing mid-copy handed the game a report that was half the
previous frame and half the next. For a button that is a one-tick glitch; for
motion it is a spike in angular velocity, which an integrator turns into aim
movement. `PadShm` gains an `input_gen` seqlock (v2.3, carved from reserved
space inside the v2 legacy region): the host takes it odd, fences, writes the 64
bytes, and stores it even; the driver samples it either side of its read and
retries once. The old code's own comment called this out as a known residual —
it is now closed rather than documented.
Version posture matches the ring's, with one simplification: no capability stamp
is needed, because an old host never writes the field and a constant 0 is
indistinguishable from "no write in flight", so a new driver against an old host
behaves exactly as it does today, and an old driver ignores the field entirely.
The Steam Deck write path had neither the seqlock nor even the trailing Release
its DualSense sibling carried; all three Windows backends now publish through
one `publish_input`.
G15 — motion-cadence observability. The host already computed the measurement a
"gyro feels floaty" report needs (client inter-arrival percentiles), but kept
ONE global accumulator, so two motion-capable pads in a session interleaved into
each other's gaps and produced a number describing neither. It also sat at
`debug` behind a `tracing::enabled!` check, so a field log arrived with nothing
in it and the only way to get the measurement was to ask for a re-run.
Now per-pad and always on, summarized at `info` when the session ends — the
moment a field report is being written. It costs one subtraction and one array
increment per sample: percentiles come from a fixed log2 histogram instead of a
growing sorted Vec, so there is no allocation, no per-window sort, and no way
for a client streaming as fast as the link allows to make the instrument
expensive. Percentiles are reported as bucket upper bounds (`_le`), which is a
factor-of-two answer to a question whose answers are orders of magnitude apart.
Gaps of 500 ms or more are counted as stalls rather than folded into the
percentiles — an interruption is not a cadence, and averaging it in would report
a healthy feed as a terrible one.
Gates. Windows CI runner .133, the drivers workspace on the real WDK: cargo
build, clippy -D warnings (which enforces the unsafe-audit lints), and fmt —
all green, against a source whose SHA-256 matches this commit's. Linux CI image:
fmt, build, clippy --all-targets -D warnings over pf-inject / punktfunk-core /
punktfunk-probe / pf-client-core / pf-driver-proto / punktfunk-host, and the
test suites including the 5 new motion-cadence tests — all green.
Not measured on glass. G6's stated gate is a sensor-rate reading (SDL
testcontroller or Steam's calibration screen) that matches the client's send
rate; that is still owed, and a driver change only a compile has seen deserves
it before anyone trusts the number.
The rows sat hard against the pinned title — the menu is the one form screen
with no subtitle, so the list started at the very top of the content band. It
now wears Add Host's explainer, which both says what the menu is for and is the
air that keeps the first row off the title. A pinned card explains what unpinning
does and does not touch, the same wording the Android dialog uses.
Caught by the screenshot pass, not by a test.
Observed in the first real session on the substrate: the loopback ran on
the minted 'Punktfunk Speakers' (silent on the host by construction), but
have_silent name-matches only the Streaming Microphone — so the capture
open logged 'desktop audio will also play on the host' (false) and
re-attempted the Steam-pair install it doesn't need. The minted sink is
recognized by id; its name honestly says Speakers, which the name rule
must keep refusing for FOREIGN instances.
Phase 1 of the gyro program (design/gyro-program.md, G1-G5) — the five
correctness fixes under it. Gyro aim integrates angular velocity over time, so
each of these is not a cosmetic wrongness: a wrong scale is every rotation being
the wrong size, a wrong clock is every rotation being integrated against a
fictional dt, and a stale sample is rotation that never happened.
G1 — the DualShock 4 calibration blob. A Sony pad does not assume a motion
scale, it reads one out of a fixed calibration feature report. Ours declared
0.5 LSB per °/s and 8192 LSB/g while the wire delivers 20 and 10000, so every
DS4-type session decoded gyro 40× too fast and acceleration 1.22× hot — since
the backend shipped. The blob now states the wire's own units (the DualSense
blob's numbers, deliberately: both pads consume the identical wire sample). Its
interleaved per-axis order is NOT a bug and stays: the virtual pad declares
BUS_USB, where interleaved is the correct layout; grouped is Bluetooth's.
The same blob lives a second time in the UMDF driver, which is a separate WDK
workspace that cannot depend on pf-inject — one wrong table in two files, where
fixing one reads as fixing it. Both are fixed, and the DS4 feature reports now
live in dualshock4_proto beside the DualSense's rather than in the Linux
backend, so there is one canonical copy to point at.
Field hosts keep the old blob until they update the host package.
G2 — the gate that would have caught it. Nothing pinned any backend's
declaration against the wire, so tests/motion_contract.rs now applies the
CONSUMER's arithmetic (the kernel's, and SDL's, which differ) to each backend
and asserts the result lands back on the wire constants — for the DualSense and
DS4 blobs, and for the Deck and Switch Pro rescales. It also parses the driver's
Rust source and re-derives the units from THAT, so the two copies cannot drift.
Verified non-vacuous both ways: re-introducing the old blob fails with "declares
a fractional 32/64 LSB per °/s", and reverting only the driver's copy fails with
"the UMDF driver's DS4_FEATURE_CALIBRATION has drifted from pf-inject's".
The wire units themselves move to punktfunk_core::input::gamepad, referenced by
the client's capture scale, the Deck/Switch rescales, and the probe — whose
at-rest vector said 16384 (a driver's number, not the wire's) and now says 1 g.
G3 — real sensor clocks. The DualSense advanced its sensor timestamp by +1 raw
unit per report (0.33 µs — a frozen clock) and the DS4 by a flat +188 (~1 ms)
regardless of the real 4-8 ms cadence. Anything integrating rate × dt off that
field got nonsense. All four backends now stamp elapsed monotonic time in their
own units via a shared SensorClock, anchored to the pad's first report so an
irregular publish loop cannot make it drift, and truncated to the field width —
which reproduces the wrap real hardware does.
G4 — motion is level-triggered and had no watchdog. merge_frame preserves the
last sample and the heartbeat re-emits it, so a feed that stops leaves the pad
rotating forever — and with G3's honest clock, at a dt that keeps growing.
Rumble and the pen plane each have an idle timeout; motion now has one too, at
100 ms. Angular velocity only: acceleration is kept, because gravity is
legitimately persistent and blanking it reads as free-fall. The SDL client
parks its gyro at zero when a slot closes, which is the case we can flush
rather than wait out. (The Apple half of this rides in PR #88.)
G5 — a pad returning inside the 300 ms replug grace keeps the same device and
skips the create path, so a different controller inherits the previous one's
touch contact and rotation — and a pad with no gyro never sends a sample to
correct it. sweep() now reports re-claims separately from drops, and the manager
clears the rich plane on one. Rich fields only: rumble and hidout dedup
deliberately survive a removal.
Gates (Linux, CI image): fmt, build, clippy --all-targets -D warnings over
pf-inject/punktfunk-core/punktfunk-probe/pf-client-core, and the test suites —
110 pf-inject unit + 6 contract + 29 pf-client-core gamepad, all green.
Not yet verified on glass; the on-glass sign/scale session is G16.
Field report from an Intel Arc + NVIDIA laptop: pinning the Vulkan rung on the Arc
iGPU silently produced D3D11VA, and there was no way to tell whether the build had
tried at all. That ambiguity was ours, in three places.
The "unavailable" log printed three of the FIVE conjuncts that gate Vulkan Video.
A device with 1.3, the features and a decode queue family — but no codec extension
— logged dev_is_13=true features_ok=true decode_family=true next to the word
"unavailable" and named nothing actionable. It now prints all five, plus which
base extensions are missing, which codec extensions are present, the decode
family's own advertised codec operations, and the device name and vendor. It also
no longer says "VAAPI/software" on Windows, where the rung below is D3D11VA.
The native-vulkan PIN refusal logged `video_decode` alone. On a device that
decodes something but not THIS codec, that reads as a contradiction: refused, yet
video_decode=true. It now carries the caps mask and the codec bit that was wanted,
so "your GPU can't" is distinguishable from "we asked for the wrong thing" — only
the second is our bug.
And `--probe-decode` is new: per-adapter Vulkan Video capability with no session,
no surface and no logical device. For each GPU it answers usable yes/no, the
driver's own decode ops, the extensions, and — when the answer is no — which
conjunct failed, in words. Separate from --list-adapters, which the desktop shells
parse line-by-line for their GPU picker and which therefore keeps printing bare
names.
The listing is ordered like pick_device (discrete first) and marks entry 0 as the
default presenter, because that ordering is very likely the reporter's actual
answer: pick_device ranks DISCRETE_GPU above INTEGRATED_GPU, Vulkan Video decodes
on the PRESENTER's device by design (that is what makes it zero-copy), and
PUNKTFUNK_DECODER does not move the presenter. So on a hybrid laptop, pinning the
decoder while the dGPU presents probes the wrong GPU entirely —
PUNKTFUNK_VK_DEVICE=<index> is the knob that moves it, and the index printed is
that value.
To keep the probe honest, VIDEO_BASE and VIDEO_CODECS moved to module scope and
the five-way AND became video_decode_gate(), called by both the probe and device
creation. A probe holding its own copy of the rule is one that eventually reports
a capability the session then refuses — which reads to everyone as a decoder bug
rather than a probe bug.
Gates: fmt clean; clippy -D warnings over punktfunk-client-session and
pf-presenter. The Linux container was unavailable (the host's disk filled and took
the docker daemon with it), so this ran on the macOS host target only — the
container leg is owed, and CI covers it on the PR.
Two gaps, both found on the shared Linux/Windows console UI.
**The settings tabs only moved for a gamepad.** They were bound to the shoulder
buttons and to PgUp/PgDn, and the legend spells PgUp/PgDn out only when NO pad is
attached — so with a controller plugged in a keyboard user had nothing to find,
and a mouse or a touchscreen could not change section at all.
The root cause was wider than the strip: `SkiaOverlay::handle_event` matched only
`KeyDown` and `TextInput`, so every mouse button, wheel and touch contact fell
past the console into the run loop, which routes pointer input exclusively at
`stream.capture` — `None` while you are browsing. Nothing in the console had ever
been clickable. Making just the pills answer would not have helped either: the
settings screen is opened with X from home, so a mouse could not reach it.
So the console gets a real pointer path:
- `Overlay::handle_pointer` carries mouse/touch in SWAPCHAIN PIXELS. The run loop
converts (it owns the window, hence the display scale, and mouse coordinates are
logical while fingers are normalised); the console then hit-tests the very rects
it drew last frame. Only DIRECT touch devices are offered — an indirect trackpad
already drives the mouse.
- Widgets act on the PRESS, not the release. The list and both carousels scroll the
focused item toward the centre, so what you pressed has slid out from under your
finger by the time it lifts; press-to-act has no such race and there is no drag
gesture to compete with.
- The hint bar became the pointer's button bar. It is already the console's only
on-screen statement of what the face buttons do, and a pointer has none — so its
Confirm/Back/Secondary/Tertiary pills are clickable on every screen, which is what
puts Settings and Library within reach of a mouse at all.
- Tab / Shift+Tab change section; PgUp/PgDn still do, and the keyboard legend now
reads "Tab".
- Right-click is Back everywhere, EXCEPT at the root: B there quits the launcher and
a right-click is far easier to fire by accident. Quitting stays explicit.
**Host cards had no menu.** Every other client hangs Wake / Copy link / Edit /
Forget off a host card; the console could add a host and connect to one, and that
was all — so a renamed machine or a fat-fingered address stayed wrong forever
unless you opened a desktop shell. UP on a saved tile now opens that host's menu,
the same gesture the Android console uses, on the one direction a horizontal
carousel leaves free.
- `ConsoleCmd::UpdateHost` edits the stored host IN PLACE. Removing and re-adding
would silently drop the fingerprint, the learned MAC, the pinned cards and the
profile binding — that is a rename, not a re-pair.
- `ConsoleCmd::ForgetHost` drops it; if it is still advertising it returns as a
discovered, unpaired row, which is the honest state.
- Forget arms on the first press and fires on the second. The other clients forget
outright; a console is driven by a thumbstick from across a room.
- A pinned profile card offers only Unpin. It is a shortcut, not a second host, and
offering to forget the host from it would blur exactly the distinction a pin draws.
- "Edit…" REPLACES the menu on the stack rather than stacking over it, so Back from
the editor doesn't land on a menu describing the host as it was before the edit.
Verified in the pf-lxcheck2 container (this crate compiles to nothing on macOS —
a bare `cargo check` there is vacuous): plain build and `clippy --all-targets`
clean under `-D warnings`, 72 tests pass. Seven are new, and cover the reported
bug directly — a press on a pill selects that tab, and each tab still keeps its
own cursor when a pointer is what switched it.
Measured on the target box: the pump wired 2 s before the provisioning
worker latched, took the cable as its write target, and the next wiring
pass would then have paired the default recording with the minted
microphone — which nothing writes into: dead mic-air until a pump reopen.
resolve_target now provisions synchronously (instant once latched; the
opt-out env is honoured), so the pump's held device and the plan's verdict
can never disagree.
An Audio wiring card (Windows hosts) below the status tiles: a readiness
badge (Ready / No microphone / No game audio / Not wired), the friendly
names carrying each role, and the degradation notes that were previously
visible only in the host log — mic withheld for game audio, the known-
degraded last resort, a narrowing endpoint. api/openapi.json regenerated
from the host build (AudioWiring + RuntimeStatus.audio); en+de messages.
RuntimeStatus gains an 'audio' object (Windows hosts): readiness
(full/audio_only/mic_only/none), the friendly names carrying each role,
and the three degradation flags (mic_withheld, last_resort, narrowing) —
the verdicts that previously lived only in tracing logs. Snapshot of the
last wiring pass (the mic pump wires at host start and on every reopen);
a status poll never triggers COM work or IPolicyConfig writes.
Opt-in "Gyro from this phone" (gyro_on_phone, off by default): this
device's IMU sources wire pad 0's motion while that pad is a controller
with no motion source of its own. On Android that gate is exact — the only
pads that forward motion are the capture links (USB DualSense / SC2,
claimed as ExternalPads), so the mirror stands down per sample whenever
GamepadRouter.padHasOwnMotion(0) says a capture link holds the index, and
sends nothing while pad 0 has no slot at all (motion never creates a host
pad). "Rumble on this phone"'s sibling, data flowing the other way: same
read-once-at-attach settings plumbing, same hardware-gated rows in the
touch and controller settings (a TV box has no gyroscope to mirror from).
DeviceGyro registers TYPE_GYROSCOPE + TYPE_ACCELEROMETER at ~200 Hz on a
dedicated HandlerThread with batching disabled (maxReportLatencyUs = 0 —
batching is poison for gyro aim), converts with the wire contract shared
with pf-client-core (rad/s → 20 LSB/°·s, m/s² → g → 10000 LSB/g; Android's
accelerometer already reads specific force, the DualSense report's own
convention), and rotates each sample from the natural-portrait sensor
frame into the controller frame by display rotation — a phone clipped
landscape yaws when the player yaws instead of rolling. The remap matrix
and unit constants are pinned by DeviceGyroTest.
A stand-down edge (capture link claims pad 0, or session teardown) sends
one zero-gyro sample so the host's virtual pad never keeps integrating an
angular velocity this device stopped producing — the gyro sweep's
stale-rotation latch, avoided by construction here.
Opt-in "Gyro from this device" (DefaultsKey.gyroFromDevice, off by default,
iOS only): while player 1's forwarded controller reports no rotation rate of
its own — no GCMotion, or the gravity-only motion an Xbox pad exposes — this
device's IMU sources pad 0's wire motion instead. The rumble-on-device
mirror's sibling, data flowing the other way: same session-scoped
UserDefaults read, same hardware-gated settings rows, same pad-0 rule.
DeviceGyro wraps CMDeviceMotion at the ~100 Hz CoreMotion ceiling on a
dedicated serial queue (not main — the controller path's main-queue delivery
is a known jitter source), converts with the shared GamepadWire constants,
and rotates each sample from the device's portrait frame into the controller
frame by interface orientation, so a phone clipped landscape yaws when the
player yaws instead of rolling. The remap matrix is derived and pinned by
DeviceGyroRemapTests.
GamepadCapture owns engage/stand-down (reconcile, suspend/resume, stop), and
suppresses pad 0's controller-motion forwarding while the mirror runs — two
writers on one pad's motion state would fight, and the accel-only stream
would stomp the mirror's gyro with zeros.
Also fixes the stale-motion latch from the gyro sweep on the controller
path: flush now parks motion at zero (keeping the last accel, so gravity
doesn't become free-fall), and the mirror's stop sends the same closing
zero. The host holds motion as state and re-emits it — a nonzero angular
velocity left behind read as endless rotation for as long as an overlay
(Control Center pull-down) kept the app inactive.
The capture-direction lookup built its endpoint id with the RENDER prefix
{0.0.0.00000000}., but WASAPI's enumeration returns capture ids as
{0.0.1.00000000}.{guid} — so the minted microphone's capture side never
string-matched the enumeration and the wiring plan paired no recording
device (audio-probe plan on the target box: mic_capture = '-'). Measured;
IMMDeviceEnumerator::GetDevice tolerated the wrong prefix, which is why
the S3 spike's direct open still passed.
A fresh CLI process has no startup worker to have finished, so the plan
devtest raced its own background provisioning thread and printed the name
ladder instead of tier-0. ensure_blocking() re-resolves existing marker
devnodes in milliseconds before the wiring pass runs.
The other half of the audio-substrate decision (spikes S2+S3 green, minted
endpoints landed in the previous commit): stop bundling a third-party
kernel driver the host no longer needs.
installer the VB-CABLE task, payload, silent-install run and the
donationware notice are gone; a suppressible notice tells
a Steam-less box that audio needs Steam INSTALLED (never
running) and that installing it later just works. A cable
from an older install is still deliberately not removed.
packer + CI -VbCableDir/VBCABLE_DIR, the staged-payload check and the
runner provisioning download are gone; SBOM drops the
redistributed-driver component.
winget the VB-Audio bundling-grant agreement becomes the honest
Steam requirement (surfaced on the unattended path where
no wizard is on screen).
docs windows-host/uninstall/security/echo say what actually
ships: no kernel-mode driver of our own, endpoints minted
from Valve's vendor-signed drivers, VB-CABLE mentioned
only as the historical fallback that keeps working.
host wording the mic-open guidance and module headers lead with Steam;
the NAME ladder itself is untouched — demoting 'cable
input' was considered and rejected (on a box where minting
transiently fails, the SSM would outrank an installed
cable, steal the silent sink, and make audio host-audible).
Reconciling with #85 (FFmpeg is gone from the client). Two of my claims
were true against the pre-merge tree and false against this one.
Note 4 said the desktop clients need no 4:4:4 decode probe "because every
rung can display full chroma — swscale converts for the software rung".
There is no swscale any more. The CPU floor is openh264 + rav1d, it is
4:2:0 8-bit by contract and has no HEVC at all, so it refuses a 4:4:4
stream rather than converting one. The client still advertises the bit
unprobed, which was the point of the original fix, but the honest reason
is different: full chroma is a hardware path (Vulkan RExt, NVIDIA today),
and what catches a box whose hardware 4:4:4 fails is note 2's codec
reconnect, not a downgraded picture. Note 13 repeated the same wrong
premise and now names both halves of its warning.
The C ABI is 17, not the 14 I read before the merge.
Textual side of the conflict: #85 rewrote the Codecs column and notes 1-3
of the same table while leaving note 4's stale 4:4:4 text alone. Theirs
kept in full; only the 4:4:4 column and note 4 are mine.
The audio-substrate program's Phase 2 (spikes S2+S3 measured green on the
target box): the host mints its OWN instances of Valve's streaming-audio
drivers and wires by IDENTITY instead of borrowing Steam's primaries —
minted.rs the provider: one devnode per role ('Punktfunk Speakers'
from SteamStreamingSpeakers.inf, 'Punktfunk Microphone'
from SteamStreamingMicrophone.inf), marker-matched across
restarts (PunktfunkAudioRole in Device Parameters — names
are NOT identity, a minted instance is name-identical to
the primaries), provisioned on a startup worker like pad
audio, retried with a 60 s cool-down from wiring passes,
defaults restored when a fresh endpoint grabs them.
wiring_plan MintedIds tier-0: the mic takes its minted device outright
(capture side paired by the provider's id — a name search
cannot tell it from the primary), the loopback prefers the
minted sink at the head of the silent tier, an operator
override still beats everything, a narrowing minted sink
demotes below real hardware, and stale ids fall back to
the ladder unchanged. Plus AudioReadiness — the
full/audio-only/mic-only/nothing classification, logged
with every plan change (§C4's seed).
audio-probe 'mint' runs the provider synchronously; 'plan' prints one
real wiring pass + readiness — the field-triage command.
Without Steam's drivers nothing changes: provisioning degrades to absent
ids and the plan keeps the name-based ladder (primaries → cable → real
hardware) exactly as before.
The Windows host job died in its clippy step: eight items in pf-encode's
split-encode policy (`SPLIT_AUTO`..`SPLIT_DISABLE`, `resolve_split_mode`,
`max_forced_split_mode`, `clamp_to_engines`) were reported as never used, and
`-D warnings` turns that into a build failure. Nothing about the encoder was
wrong — the items simply have no reader in one particular build of the crate,
and nothing was telling the compiler that.
`codec.rs` compiles on every platform, but the split policy only ever has a
caller on Linux (the libav NVENC path reads it unconditionally) or on Windows
with the `nvenc` feature (the direct-SDK backend). A featureless Windows build
of pf-encode has neither, so every item in the cluster is genuinely dead there.
Gate them on the union of their callers' cfgs, the way `forced_split_width`
next door already is.
The step lints pf-encode itself WITH `--features nvenc,amf-qsv,qsv`, where the
items are live, which is why this was invisible there; the failure came from the
next command in the same step, `clippy -p pf-vdisplay`, which pulls pf-encode in
as a plain default-features dependency. Same item-level `dead_code` trap this
crate has now hit five times.
Verified: default-features pf-encode reproduces all eight errors before the
change and none after (macOS default-features exercises the identical
"cluster has no caller" arm as featureless Windows — the two remaining errors
there, `vbv_frames_env` and a redundant closure call, are pre-existing and
macOS-only; both items have real Windows callers). Linux default-features and
Linux + nvenc `--all-targets` both stay clean, so the callers still see the
policy. `cargo fmt` clean.
The support matrix said the desktop clients' Full chroma switch "has no
effect today" and that only the Apple client asks for 4:4:4. Both stopped
being true in July: `clients/session/src/main.rs` advertises VIDEO_CAP_444
whenever the setting is on, deliberately with no client-side probe, because
every desktop decode rung can display full chroma — the Vulkan presenter
samples the 2-plane 4:4:4 pool formats and swscale converts for the software
rung. So Linux, Windows and Apple all ask; Android is the one that genuinely
doesn't implement it.
The other half was HDR. `9f72a3b6` gave the Windows IDD-push capturer a
packed 10-bit BT.2020 PQ RGB output, so NVENC encodes HEVC Main 4:4:4 10 and
the two compose — the matrix still said "4:4:4 and HDR together is refused",
and hdr.md still called PyroWave the only exception. Linux is the side that
keeps the trade: handshake.rs resolves the depth back to 8 for a 4:4:4
session, so full chroma wins and the stream is SDR.
Three cells move ❌ → ⚠️ rather than ✅ on purpose. The client half is
unconditional, but the host half is not: HEVC 4:4:4 means an NVIDIA host, or
PyroWave on any vendor. The notes say which, and point at the stats overlay's
`4:4:4→4:2:0` tag — this negotiation is the one that fails loudly.
Also: C ABI version 13 → 14; PyroWave's ≈8K 4:4:4 block-index ceiling now
has a note; and the roadmap no longer calls Intel 4:4:4 a hardware limit,
which the matrix and vaapi.rs both contradict — VCN can't, VAAPI hasn't.
Spot-checked and left alone as still accurate: the Linux client clipboard
stub, VAAPI declining 4:4:4, Android having no 4:4:4 at all, and the wire /
driver / gamepad-channel versions.
main moved 93 commits while this branch ran. Two conflicts, both where main's new
work sat next to M10's excision:
packaging/flatpak/io.unom.Punktfunk.yml — main added the vendored gamescope WSI
layer (the only route to HDR on a Deck) and, before it, a vulkan-headers module.
Took both: this branch predates them and deletes neither. But the headers module's
stated consumer was pf-ffvk's bindgen over FFmpeg's hwcontext_vulkan.h, and M10
deleted pf-ffvk — so it now reads as dead weight to the next person. It is not:
the WSI layer IS a Vulkan layer, compiles against those headers, and builds after
it, so module order is the dependency. Rewrote the rationale to say so, including
why dropping it would be expensive to discover — flatpak.yml has no pull_request:
trigger, so a manifest break reaches main invisibly and a tag then ships no Linux
flatpak. Also recorded that the native decoder needs nothing from there: pf-vkdecode
reaches Vulkan through ash, which is pure Rust bindings, no bindgen, no C headers.
crates/pf-console-ui/src/screens/settings.rs — main restructured the gamepad
settings into TABS, which removed the per-row section headers; this branch had left
Some("Video") untouched from the merge base and added the pre-M10 decoder migration
next to it. Git could not tell those apart. Took main's structure (no header, its
deliberate change) with this branch's migration layered on: a stored `vulkan`,
`vaapi` or `d3d11va` names no preset in the tabbed list and would render as "—",
then silently rewrite the user's preference on the next save.
Gates on the merged tree, Linux container: fmt clean; cargo check --workspace
--all-targets clean; clippy --workspace --all-targets -D warnings clean; tests
green across pf-vkdecode (187), pf-client-core (163), pf-console-ui (58) and
punktfunk-host (447 of 448 — the one failure is the pre-existing
gamestream::stream::tests::sender_delivers_batches, a UDP-loopback EINTR under
qemu that fails identically on a pristine HEAD).
The S1-S3 spikes from windows-audio-endpoints-and-vbcable.md as one
runnable devtest (no game, no client, ssh-drivable):
audio-probe ssm S3, the decision gate: mint a SECOND devnode of
Valve's Steam Streaming Microphone driver and
prove the pair end to end (tone into its render
endpoint must come back out of its capture
endpoint). Pass = a punktfunk-owned virtual mic
needs no VB-Cable wherever Steam is installed.
audio-probe sink S2: mint a Speakers instance, park the DEFAULT
playback on it, tone through the default device,
WASAPI-loopback the instance - the desktop-audio
capture path minus the game.
audio-probe sss-primary S1: the primary Speakers' known-silent loopback,
re-measured, with mix format + steam.exe state.
audio-probe cleanup remove every probe-minted devnode (marker value
in Device Parameters, never name-guessing).
pad_endpoint grows the first slice of the design's §C1 shared minting
surface: create_media_devnode(desc, hwid, mark), bind_driver(hwid, inf),
find_capture_endpoint_for_devnode — the pad provisioner now calls the
same functions. The probe restores whatever default devices the minting
disturbed before it exits.
install.rs (landed 2026-08-05 with the security-review remediation, while
the Windows CI runner was down) fails windows-host.yml's clippy gate:
#![deny(clippy::undocumented_unsafe_blocks)] wants the SAFETY comment on
the line preceding EACH unsafe block, and three blocks didn't have one —
two sat behind a comment anchored to the enclosing closure/neighbouring
statement, and EqualSid had none at all. Comments only; no behavior
change.
The wiring plan reserved the mic target unconditionally first, so on a box
without VB-Cable the mic took the Steam Streaming Microphone — the only
working client-only loopback sink — and desktop audio fell to the
known-silent Speakers last resort: a headless Steam-only host streamed
SILENCE (the 2026-08 field case), and the installer's 'optional (mic
passthrough)' wording never warned anyone.
The mic may now hold the Streaming Microphone only while the loopback still
gets a preferred (non-last-resort) pick without it — another silent sink or
real hardware. Otherwise the loopback takes the endpoint and the mic falls
to a lesser candidate or is honestly withheld (Wiring::mic_withheld), with
the open error naming the trade and the remedy. An operator
PUNKTFUNK_MIC_DEVICE override is exempt: an explicit choice may still
strand the loopback on the last resort.
Also: the Steam-pair auto-install latch is now once per INF-state instead
of once per process — an attempt made while Steam was absent re-arms when
its driver INFs later appear (files are invisible to the endpoint-set
fingerprint, so nothing else would ever retry), and a withheld mic skips
the pointless reinstall (the pair exists; the plan gave it to the loopback).
Two defects, found while tracing M8's codec-fallback reconnect and recorded
verbatim in d5e23146 as out of scope there.
A client retry re-sends Hello::launch verbatim, and the host launched
unconditionally. Steam and Epic URIs hide it — the launcher focuses the running
copy — but a gog:/custom: target really did start a SECOND COPY of the game. The
client cannot fix it by dropping the field: on Linux the per-session gamescope is
re-adopted through pf-vdisplay's display registry, whose reuse key includes the
launch command, so a retry without it orphans the running game.
And the retry minted a fresh launch_stamp, so procscan refused to adopt a game
started more than 2 s before it — the game was minutes old, so a reconnected
session had no game-exit detection for the rest of its life.
Both are now answered by a launch registry (launchreg.rs): one record per (client
fingerprint, library id), written at launch time and INDEPENDENT OF THE
TERMINATION POLICY. That independence is the point. The existing fingerprint-keyed
reclaim only exists under GameOnSessionEnd::Always — under the default Keep,
arm_grace is never called, so nothing was recorded at all in exactly the
configuration the defect was reported in.
The design correction that matters: at launch time the host knows NOTHING about
the game's processes — that is the premise of the whole lease design. So identity
flows BACKWARDS from the watcher, which publishes the concrete ProcRefs it
adopted, and the registry's liveness is Scanner::alive over that recorded set,
re-verified by (pid, start). Never a re-scan by spec: a later scan would find a
copy the player started since, and adopting that is what procscan's rule 1
forbids. The published set is never cleared on exit either — the last thing the
watcher saw is what makes a quit game read Gone rather than "no opinion", which
is how it becomes relaunchable at once instead of being suppressed for the window.
On rule 1: an adopting session inherits the older floor, so its own find() admits
what the ORIGINAL session's lease already admitted for its whole life. That is the
correct reading of "the same launch, continued" and not a new exposure — rule 1
forbids adopting processes that PREDATE the launch, and these postdate it.
The match rule is pure and total (covers()): liveness is authoritative where it
has an opinion, and only Unknown falls through to the tie-breakers — a live holder,
or a 90 s in-flight window for a re-dial while the launcher is still working. Gone
beats both, deliberately: a title that crashed on startup must relaunch at once.
Both race orders are handled and neither is relied on. Teardown-first takes the
Running arm; handshake-first (a fast re-dial on a half-open connection) takes the
holders>0 arm, and the old teardown then sees superseded() and does nothing —
without which, under Always, it would arm a grace the new session had already
passed its chance to reprieve, and the reaper would kill the new session's game.
Two tradeoffs taken deliberately: a custom: command with no detection hints stays
Unknown forever, so that reconnect trades game-exit detection for not
double-spawning; and IN_FLIGHT_WINDOW is a fixed 90 s rather than sharing
disconnect_grace_seconds, because the two have opposite failure costs — grace
being wrong leaves a game running, this being wrong silently swallows a launch the
player asked for.
Gates: fmt clean; clippy -p punktfunk-host --all-targets -D warnings green in the
Linux container; 418 passed, +9 exactly the new tests. One failure,
gamestream::stream::tests::sender_delivers_batches, is pre-existing and
environmental — a UDP-loopback EINTR under qemu at stream.rs:1697, outside every
hunk in this change (the last is at +448), and it fails identically on a pristine
HEAD. I reproduced both the failure and its location myself rather than taking it
on report.
⚠ OWED: the Windows leg is COMPILE-UNVERIFIED. cargo check --target
x86_64-pc-windows-msvc dies in ring's C build on macOS and xcheck.sh does not
cover punktfunk-host. The Windows edits are small restructures of existing
branches plus a bool assignment, reasoned through but seen by no compiler. Run it
on .133 before this merges.
I narrowed that exposure by inspection afterwards, and it is smaller than the
blanket warning suggests. The change presents exactly two things to a Windows
compiler that a Linux one did not already see. launchreg gates only alive_count
(lines 227/231), whose cfg(any(linux, windows)) arm calls
Scanner::system().alive(procs) — the identical call gamelease.rs:563 already makes
in code that compiles on Windows today. And the Windows launch arm at
native/stream.rs:1666 reads only ungated bindings the Linux arm type-checks thirty
lines below it (adopt_launch:1658, spawned_now:1663, launch_claim:1463) and calls
only the pre-existing library::launch_title. No new type, no new signature, no
Windows-only API.
That is an argument, not a compile. The run on .133 is still owed.
clippy's undocumented_unsafe_blocks (deny) flagged the three blocks that
81039581 introduced: the SAFETY comment sat outside the closure, so
IsValidSid/EqualSid inside it read as undocumented, and from_raw_parts
shared a comment that only covered the GetLengthSid line above it. Windows
host clippy is the only leg that lints this cfg(windows) code, red since.
The LAN-registry docker login only serves the Push step (Reconcile and
Tag-for-release authenticate via curl -u), but it ran unguarded — so a
hit=true leg landing on a host with a misconfigured docker daemon failed at
login with nothing to push (run 16044/16013 f44 leg). Gate it like Build/Push.
docker image prune -af --filter until=2h keyed on image CREATION time, so a
base image built days ago that merely had no container at that instant was
"aged" — including one a job had just pulled and not yet created. Measured
2026-08-07: three job failures, each coinciding with a prune tick to the
second ("No such image: …punktfunk-rust-ci:latest", every step cancelled),
plus a 4-7 GB re-pull of every idle base image within minutes.
The routine tick now retires only what this host actually accretes — per-SHA
app tags older than 2h (their creation time IS the local build time) — then
sweeps dangling layers, which cannot touch a tagged image. The blanket -a
prune survives only in the near-ENOSPC burst guard, where one re-pull beats
every concurrent job dying.
docker-reclaim.{sh,service,timer} are the hourly leak reclaimer that so far
lived hand-installed on home-runner-1 only; home-runner-2 went without it and
re-accumulated 176 leaked volumes (~60 GB) until jobs died of ENOSPC on
2026-08-06/07. Checked in so both hosts install the same files from here.
The AV1 use-after-free fix (cdd1f3ef) stabilised the wrong half. NVIDIA was
measured retaining pColorConfig, so StoredParamsAv1 boxed the colour and timing
blocks — but OwnedStdAv1SequenceHeader kept the Std struct ITSELF inline, so the
pStdSequenceHeader we handed vkCreateVideoSessionParametersKHR was a stack
address inside ensure_parameters, dead the moment it returned. The fix worked
because of WHICH pointer that driver happened to hold. A driver retaining the
outer one instead — no more of a spec violation than retaining pColorConfig was —
reproduces the original bug exactly: plausible pictures, wrong content, no error
and no counter moved.
The same shape was in the shipping codecs, one step further from evidence: the
H.264 and H.265 create paths pointed pStdSPSs/pStdPPSs/pStdVPSs at function-local
Vecs, and both Add paths handed over the wrapper's inline std field and then moved
the wrapper. Those are spec-legal — the object stores copies — and have never
misbehaved on the fleet. They are fixed anyway, because that is precisely what was
true of H.264/H.265 before the same class of bug was found in them, and a
correctness argument that reduces to which vendor we tested is not one.
So: the Std struct is boxed inside each owning wrapper (one level out from what
_color_backing already did), and the contiguous create-time arrays are now fields
of the stored parameters, assembled at their final address. Identical bytes at
identical offsets — only where they live changed.
The line drawn deliberately, in prose at session.rs:29: Std DATA is pinned; the
VkVideoSessionParametersCreateInfoKHR chain itself is not. Retention there would
be a different and far more extreme class of driver bug, and pinning it needs a
self-referential struct over lifetime-parameterised builders.
⚠ NOT hardware-verified. No GPU has run this — the fleet is unreachable and the
250/250 parity that proved this code bit-exact cannot be re-run. That is why the
change is constrained to address stability alone, and why it ships five CPU-only
tests instead: three capture the pointer handed to Vulkan, perform the real move,
and assert it survives — each verified FAILING first, with genuinely differing
addresses, not a tautology. Two more pin the create-array ownership; those fail
before the fix as compile errors rather than assertions, because the pre-fix bug
there is a dangling pointer and asserting on it is UB.
Also: caps.rs claimed the borrow checker pins a profile chain between wire() and
its last use. False at exactly one site — decoder.rs took a raw *const, ending the
borrow, leaving nothing but inspection to stop a future editor moving the chain
before create_query_pool. Correct today, guarded by prose, which is how the first
bug shipped. It is now compiler-enforced: the pointer write and the create call
live inside one helper that takes the profile by reference, so the borrow is held
across both by the signature. An audit cleared the chains otherwise — no entry
point we pass one to retains it.
Gates: fmt clean; clippy -D warnings over pf-vkdecode AND pf-client-core in the
Linux container (its only real consumer, which cannot build on macOS at all —
wol.rs uses deps its manifest gates to linux/windows, so workspace clippy has
never passed there and does not now); 187 lib tests green on Linux, up from 182.
The last coverage gap, and only worth building once S1 proved it possible: the
Windows backend drives NV_ENC_DEVICE_TYPE_DIRECTX, and an in-place splitEncodeMode
change had never been tested there. It works (071358cb), so the arbiter is now
ungated from Linux-only to the union of both direct-SDK backends and wired into
windows/nvenc.rs: the submit stamp, the feed hook on AU completion,
apply_split_mode, split_key, arm_split_arbiter, and set_send_spread_us.
Same gates as Linux, and they are correctness conditions rather than preferences:
opt-in while it earns trust, an operator PUNKTFUNK_SPLIT_ENCODE pin always wins,
a cached verdict short-circuits, >=2 engines, never H.264, and the sub-frame
trade is only entered when the host has actually reported a send spread to price
it with. The one Windows-specific difference is that `async_rt` is a real
possibility here (opt-in two-thread retrieve) and the arbiter refuses it, because
under pipelined retrieve the submit->AU span includes queue depth and the
comparison would be noise.
⚠ Two more instances of the same item-level dead_code trap, caught by the Windows
run and not by reasoning -- that is now 4 and 5:
- `clear_split_verdicts` is called only by the Linux on-hw test, so it is dead on
Windows; gated to `all(test, target_os = "linux")`.
- The arbiter methods first landed inside `impl Encoder` rather than the inherent
impl (the anchor I used, supports_chunked_poll, is a trait method), which the
compiler caught as "not a member of trait Encoder".
Verified .158 (RTX 4090 / Ada, driver 610.88, D3D11): clippy --features nvenc
--all-targets -D warnings clean, and 2 on-hardware NVENC tests green including S1
re-run with the arbitration code in place (engines=2 latched, DISABLE->TWO_FORCED
accepted, zero IDRs, reverse accepted). Verified .21: clippy clean with AND
without the nvenc feature, 65 unit tests, 25/25 NVENC on-hardware. fmt clean.
The plan's M10 checklist named "the about.toml carve-out that puts FFmpeg
outside the automated licence gate". There is no such stanza — I looked, on this
branch and on origin/main. The carve-out is structural, which is worse: cargo-about
walks the CARGO graph, so a native library reached through a permissively-licensed
-sys crate is invisible to it. ffmpeg-sys-next is WTFPL and passes the gate
cleanly while the LGPL libavcodec it link-imports is never harvested at all.
So about.toml's own claim to be "exactly the regression guard we want against a
copyleft dependency silently entering the linked set" was overstated: it did not
catch FFmpeg entering and would not catch the next one. The comment now says so,
and says where the LGPL obligations are actually discharged instead.
The one genuinely good piece of news is recorded too: since M10 the client links
no FFmpeg, so for every client artifact the crate graph and the linked set
coincide and the gate finally means what it appears to mean. The gap is the
host's alone.
Gate: cargo about generate about.hbs --fail — passes.
On a Mac with brew's opus installed, audiopus_sys found it via pkg-config and
statically linked it into the aarch64 slice — a lib built for the RUNNING
macOS (minos 26.0, tripping the script's own version guard) and existing only
for the host arch, so the x86_64 slice silently fell back to the vendored
build and the two slices shipped different libopus builds. Force the vendored
CMake build for every slice (OPUS_NO_PKG_CONFIG=1), with the CMake policy
floor modern CMake (>=4) needs to accept libopus's old cmake_minimum_required.
The shared JitterPolicy grew an adaptive target floor — clustered genuine
underruns raise the live target a step at a time up to max_target_ms, a long
quiet spell relaxes it back — and the three Rust rings all run it via
note_read. The Apple ring is the one hand-written mirror, and it mirrored the
shed half but not the growth half: its target was pinned at the 20 ms base
forever. On Wi-Fi that bunches arrivals (power-save is the classic; the field
MacBook report is the symptom), 20 ms is regularly shorter than one delivery
stall, so the ring re-primed through every stall for the whole session —
crackle that never got better, on exactly the client where a Moonlight with a
deeper buffer sounds fine on the same host and network.
The ring now carries the full mirror of note_read: 3 underruns inside a 5 s
window grow the target 10 ms (capped at COREAUDIO's 70), 30 s of quiet gives a
step back, and the write-side hard trim follows the grown target (including
the Rust policy's target+quantum guard, which the mirror also lacked). New
tests pin the mirror to the Rust suite's expectations — growth, relax, the
cap — plus the field scenario end to end: bunched 60 ms deliveries with every
fourth burst 30 ms late converge to a silence-free tail instead of crackling
forever.
A field report: game audio on a MacBook (M1) crackles over Wi-Fi against a host
that plays clean to other clients. The Apple client is the one client whose
Opus decode lives in core (punktfunk_connection_next_audio_pcm — AudioToolbox
has no multistream path), and that decoder only ever decoded packets that
ARRIVED. The Linux, Windows and Android decode loops all feed an
AudioGapTracker and synthesize libopus packet-loss concealment for every
packet the wire lost; the in-core path had the tracker sitting unused in the
same crate. So on Apple every lost 5 ms datagram — at ~200 packets/s over
Wi-Fi, a steady trickle — landed in the playout ring as a hard time-domain
gap: a click per loss, sustained crackle under real loss. The redundant-plane
recovery (0xD2) hides single losses when the host grants it, which is exactly
why the survivors are the burstier gaps that need concealing most.
The decode now runs through the same accounting as everyone else: concealed
frames land in front of the arriving frame in one contiguous buffer (the
embedder just writes it to its ring), a DTX marker advances the accounting
without being decoded, and the output buffer is pre-sized for a full
concealment run so the borrow-until-next-call pointer can never dangle.
Unit-tested against real libopus: gaps, duplicates, DTX-after-loss, and the
50 ms cap.
Everything the split-encode programme rests on had been proven only on
Linux/CUDA. The Windows backend drives NV_ENC_DEVICE_TYPE_DIRECTX, so none of it
transferred by assumption -- and if the driver refused an in-place split change
there, Windows arbitration would simply not be buildable.
RESULT on the RTX Windows box (RTX 4090 / AD102, driver 610.88, D3D11):
engines=2, latched by query_caps (WP1.1's probe, validated on Windows
hardware rather than inferred from Linux)
DISABLE -> TWO_FORCED via nvEncReconfigureEncoder, resetEncoder=0: ACCEPTED,
ZERO IDRs, and the reverse likewise.
So the foundation now holds across three platform x arch x driver combinations:
Linux/CUDA Blackwell 610.57.04, Linux/CUDA Ada 610.43.03, Windows/D3D11 Ada
610.88.
⭐ UNBLOCKS ALL FUTURE WINDOWS ON-HARDWARE TESTING. pf-encode's nvenc test
binaries were believed unlinkable on Windows ("NvEncodeAPICreateInstance
unresolved", recorded as pre-existing and worked around by only ever running
clippy there). They link fine given the SDK import library:
RUSTFLAGS='-L native=C:\Users\Public\nvenc -l nvencodeapi'
`-L` alone is not enough -- without a `-l` nothing pulls the archive in, which is
why the earlier attempt still failed. ⚠ This is TEST-BINARY-LOCAL and must stay
that way: production deliberately dlopens NVENC rather than link-loading it, and
an unconditional link-load is the known crash class on non-NVIDIA Windows hosts.
⚠ Box note: the RTX Windows box answers on .158, not the .173 in its memory
entry, and `Administrator@` there resets the connection right after
SSH2_MSG_SERVICE_ACCEPT in a way that reads like the host being down -- the
working login is "Enrico Bühler"@192.168.1.158.
windows-host.yml called FFMPEG_DIR "the same BtbN lgpl-shared x64 tree the
Windows CLIENT links against". Since M10 the client links no libav* at all and
windows.yml sets no FFMPEG_DIR, so the sentence pointed a reader at a link that
does not exist. The provisioning script still fetches the tree — for the host
alone — which is the part worth saying out loud, because the next person to read
it will wonder why a client-provisioning step still mentions FFmpeg.
The verification gap flagged in 01294e3a was real. `.133` came back up and the
WP4 commit failed Windows clippy: `forced_split_width` is used only by the libav
NVENC path (`enc/linux/mod.rs`), but it was added to `codec.rs`, which compiles
everywhere -- so it is dead code on Windows and `-D warnings` rejects it.
Third time this crate has hit the same item-level dead_code trap (see
`subframe_env_forced`, and the arbiter items in `nvenc_core`), and the third time
it was caught by actually running the Windows check rather than by reasoning
about it. The comment on the gate says so, since the pattern is clearly not
self-evident from the code.
Verified .21: clippy -D warnings clean both WITH and WITHOUT the nvenc feature,
65 unit tests. Verified .133: Windows clippy --features nvenc --all-targets
-D warnings clean, zero errors, zero dead_code. fmt clean.
cargo tree -p punktfunk-client-session finds no ffmpeg. The host still does,
which is the whole point: pf-encode keeps libavcodec unconditionally and no
host workflow, packaging script or licence file was touched.
Deleted: crates/pf-ffvk, video_vulkan.rs, video_vaapi.rs, video_libav.rs, the
libavcodec half of video_d3d11.rs, the av_log machinery, ffmpeg::codec::Id as
the decoder's vocabulary (the quic CODEC_* wire constants now serve, which is
why the evidence table was keyed on them), DecodedImage::VkFrame and ::Dmabuf,
the presenter's AVVkFrame lane, and the ffmpeg-fallback feature with
everything behind it. DrmFrameGuard collapses from an enum to a newtype, which
removes an unsafe impl Send. Roughly 25,000 lines.
Then the CI, packaging, licensing and docs work the plan's §6 lists: the
Windows workflows lose FFMPEG_DIR, PF_FFVK_VULKAN_INCLUDE and their PATH
prepend; the MSIX loses its DLL wildcard; the client .deb stops emitting libav
sonames on its own because depends come from dpkg-shlibdeps; arch, flatpak and
nix drop the dependency; and the README's "FFmpeg 7 or 8" contract narrows to
the host.
Three defects reached users' machines in the first cut, and none was in the
deletion itself.
All three desktop Settings UIs offer vulkan, vaapi and d3d11va as stored
decoder values, so those strings sit in shipped settings files today. Refusing
them by name — which is the correct rule for a stale pin — would have bricked
every upgraded client whose owner ever touched that dropdown. They now migrate
onto the native rung for the same hardware family, at decoder construction AND
at each dialog's lookup, because a legacy value that matches no preset
displays as "Automatic" and silently rewrites the user's preference on the
next save.
M9's evidence filter was deleted on the argument that with no libavcodec twin
below, barring an unproven rung removes hardware decode rather than moving
down one rung. That is true on Windows and false on Linux for Intel and every
unknown vendor id, where prefer_vulkan_first is false and the order is
native-vaapi → native-vk: a rung that has decoded nothing anywhere sitting
above one that is 250/250 on three drivers. Every Intel Linux desktop would
have moved from libavcodec VAAPI, shipping for years, onto pf-vaadec by
default — and a rung that constructs and then produces wrong pixels leaves
only by the error-streak demotion, which this codebase already documents as
not tripping on the B580's strobing. The filter is restored as a narrow, pure,
testable rule: an unproven rung yields to a proven one, and to nothing else.
Windows deliberately passes no rung below, because that vendor family is the
one with a measured wrong-pixel report against Vulkan decode, and trading no
evidence for evidence of corruption is the wrong direction.
And the notices still said FFmpeg was bundled. The root file is what both
desktop clients include_str! and what the MSIX ships, three lines under the
new card saying no FFmpeg is bundled; Apple's Acknowledgements said it too, on
iOS, tvOS and macOS. The generator now emits four per-client files scoped by
transitive closure — 0 FFmpeg mentions in each, verified — while the root file
keeps it for the host. That also ends the standing false attribution of
ffmpeg-next, GTK4, windows-rs and the NVENC SDK to an iPhone.
Windows has no reachable box, so it was compiled instead: a cross clippy at
-D warnings on x86_64 and aarch64-pc-windows-msvc with the C toolchain stubbed
so build scripts run without linking. That gate immediately caught an
include_str! path one directory too deep, which nothing else could have.
Gates: container clippy -D warnings, 160 tests, workspace check, both Windows
targets clean, client ffmpeg count 0 and host 2. The four decode crates are
untouched, so the hardware rungs' 250/250 stands.
⚠ Owed and unrun: no GPU has executed any of this milestone. M8's on-glass
software check, M7's D3D11 and VAAPI AV1 hardware legs, and M9's field bake
all still want hardware, and the bake window and criteria remain the user's.
The libav NVENC path carried its own inline copy of the split decision and had
already drifted from the direct-SDK selector: it hard-coded a 2-way split
regardless of engine count, and had no depth rule at all. That is the drift the
shared resolver was extracted to prevent, and the copy quietly reintroduced it.
Routing it through `resolve_split_mode` needed the policy to MOVE. `nvenc_core`
is gated on `feature = "nvenc"`, but the libav path is precisely the build where
that feature is OFF (`PUNKTFUNK_NVENC_DIRECT=0`, and the featureless packages --
the packaging gap this project has been bitten by before). So
resolve_split_mode / max_forced_split_mode / clamp_to_engines, plus a new
`forced_split_width`, now live in `codec.rs`, which is always compiled and
already owned SPLIT_FORCE_PIXEL_RATE.
That means the NV_ENC_SPLIT_ENCODE_MODE values had to be hand-written as plain
constants, since the SDK enum does not exist without the feature. They are
therefore pinned: `nvenc_split_constants_match_the_sdk` (feature-gated, the only
place both are visible at once) asserts all five against the real enum, so the
copies cannot rot.
⚠ Only the FORCED outcomes are actionable on the libav side -- libavcodec's
`split_encode_mode` AVOption is its own vocabulary and our DISABLE is the NVENC
enum's 15, which would be meaningless there. DISABLE/AUTO both map to "leave the
option unset", which is exactly today's behaviour (unset = the driver's auto).
`engines = 0` ("not probed") maps to 2-way, preserving what that site always did;
a 3-NVENC part gets the wider split only on the direct-SDK path, which is the one
that actually probes.
⚠⚠ VERIFICATION GAP: .133 went down mid-change (no ping), so the WINDOWS leg is
UNVERIFIED. This matters more than usual -- the Windows backend imported
resolve_split_mode from nvenc_core and that import had to move too, which a grep
caught rather than a compiler. Re-run before trusting it:
cargo clippy -p pf-encode --features nvenc --all-targets -- -D warnings
Verified .21: clippy -D warnings clean BOTH with and without the nvenc feature
(the featureless build is the whole point of the move) and with
nvenc,vulkan-encode; 65 unit tests incl. the new constant-parity test; 25/25
NVENC on-hardware; punktfunk-host clippy clean. fmt clean.
`ffmpeg-fallback` on pf-client-core, default off on the crate. With it off the
libavcodec rungs are not compiled, pf-ffvk leaves the dependency graph, and no
ladder or demotion arm names them; with it on each sits exactly where it sits
today, directly below its native twin. That is the switch which makes M10 a
deletion rather than a redesign.
The bake window and the regression criteria are the user's, per the plan, and
nothing here claims the M9 gate is met.
The hard part was not the feature, it was honesty. Two of the four native
rungs have never decoded a frame on any hardware — native VAAPI at all, and
native D3D11VA's AV1 leg — and making those the default would assert evidence
that does not exist. So admission is per rung and per codec: a pair with
hardware evidence joins `auto` always; a pair without it joins only when
nothing proven is left below it (a build with no FFmpeg twin, where the
alternative is not a proven rung but the CPU) or when the user asks with
PUNKTFUNK_NATIVE_FIRST=1. Pins bypass it, so a lab run can still reach any
rung.
The shipping default therefore changes in exactly three ways, all
evidence-backed: AV1 `auto` takes native Vulkan (250/250 bit-identical on an
RTX 5070 Ti), Windows H.264/H.265 `auto` takes native D3D11VA above its FFmpeg
twin (parity on two GPUs plus a 30-minute soak), and a failing Vulkan rung on
Windows demotes to native D3D11VA first. Everything unproven is byte-for-byte
as it was.
The evidence state is written where it cannot rot: a table in video.rs's
module docs, the same facts in code as `native_evidence()`, a test asserting
them in both feature states, and a per-session log line carrying the rung, the
codec, whether hardware has verified that pair and the evidence string — at
WARN when it has not. A support engineer reading a log can now tell proven
from assumed without asking anyone.
Termination needed a new guarantee. With the FFmpeg twins gone, two native
rungs in opposite per-vendor orders could hand a session back and forth
forever, so a rung once entered is never re-entered and the walk is monotone
to software. The never-delivered fall-through still works: with the feature on
it is unchanged, and with it off it is redundant, because the next candidate
already IS the rung below.
⚠ ffmpeg-next remains a hard dependency of pf-client-core, deliberately. What
is left off-feature is three type-level residues — the codec-id vocabulary,
the AVVkFrame guard that is pf-presenter's public import, and a pixel-format
in one signature — every one of them an M10 §6 line item. Deleting them here
would mean deleting the presenter's FFmpeg lane, 55 call sites, in a milestone
whose gates cannot run a GPU. No libavcodec decoder is opened in a default
build.
⚠ video_d3d11.rs was gated item by item rather than wholesale, and nothing in
this tree compiles it — it needs a Windows check before anyone trusts it.
Gates: both feature states, container clippy -D warnings and 158/159 tests,
workspace check. The four decode crates are untouched, so the hardware rungs'
250/250 stands.
The ladder's last rung no longer runs FFmpeg. H.264 decodes through openh264,
AV1 through rav1d, and HEVC is refused outright: no permissively licensed
software HEVC decoder exists, so an HEVC session that exhausts its hardware
rungs now tears down and re-dials advertising HEVC-less caps, and the host
picks H.264. The plan calls that a first-class path; it is one.
swscale is deleted, and with it the BT.601 default that its correction code
existed to undo. Colour on the H.264 lane now comes from the same
pf-bitstream planner every hardware rung submits from — openh264 reports no
VUI at all — and AV1's comes per-picture from the sequence header. One colour
source, one CSC: the old default is unrepresentable rather than merely fixed.
Frames reach the presenter as three tightly-packed planes through the planar
CSC pass, which had to be un-gated from the pyrowave feature and its device
probe, since the last rung must exist on devices that failed that probe.
rav1d rather than the dav1d crate, deliberately and against the plan's
literal wording: dav1d-sys is system-deps-only, so it would add a system
library and a .pc file to every client package — in the milestone family
whose excision checklist exists to delete exactly that. rav1d is the same
decoder, same licence, statically linked. The cost is honest: no-asm builds
on both decoders, and software throughput is still unmeasured.
The colour test is the milestone's exit criterion, so it is built to fail.
Three fixtures, and a mutation check: hardcoding the swscale default turns the
red bar to [255,24,0], and swapping Cb/Cr turns red to blue — a silent error
no metadata assertion could catch. Review then disproved the range half of it
numerically: with eight saturated bars, decoding the full-range fixture with
the wrong range gives max error ZERO, because a mismatch only pushes values
outside [0,1] where the shader clamps. A mid-tone was added; the wrong range
now costs 11, well past the tolerance. The exit criterion I set was
satisfiable by a test that proved nothing.
Two blocking defects, both emergent rather than local.
Software AV1 on a 10-bit stream never reached its typed refusal: rav1d is
built 8-bit-only and returns ENOPROTOOPT, which the send loop turned into a
generic error, so the pump's typed downcast missed and every AU failed
identically — a permanent freeze on precisely the shipping case, since AV1 is
advertised only where hardware AV1 exists and hardware AV1 plus HDR is Main
10. The shape is now read from the sequence header before any byte reaches
the decoder, exactly as the H.264 leg reads the active SPS.
And the new Reconnecting phase was the first state that is not streaming, not
connecting, and still holding a live stream — which opened all three guards
that had made a second launch impossible. Pressing A assigned over `stream`
where every other site shuts down first, and StreamState has no Drop, so the
old pump was detached: a second live session still submitting to a Vulkan
device that gets destroyed underneath it. Nothing about the reconnect was
wrong in isolation; the defect lived between a new state and three guards
nobody re-examined. Start is now defensive and the retry raises the
connecting modal, so the UI matches the state and B can cancel.
Also closed: retry_caps was computed, tested and never applied, so a shape
refusal could end a session reporting no codec available while a working
retry existed; the retry inherited force_software sticky-true, landing an
HEVC→H.264 fallback on software H.264 with working hardware H.264; it
re-dialled with a stale mode; the CPU present arm had no survivable-failure
handling where the pyrowave arm — same pass — has it; HEVC is no longer
advertised when the decoder is pinned to software; and the software rung now
feeds the recovery-point SEI it already had in hand to the re-anchor gate.
⚠ Two host-side gaps found while tracing, neither in scope here: Hello::launch
is NOT idempotent (gog:/custom: targets spawn a second copy on a retry; the
field is kept verbatim because dropping it orphans the gamescope display whose
reuse key includes the command), and a reconnected session can never adopt a
game predating its own launch stamp, so it has no game-exit detection.
⚠ OWED: the on-glass software run. ~200 lines of new Vulkan on a path that
only runs because the GPU already failed, and no driver has seen it. The
review's minimum check is sync validation enabled, a non-multiple-of-16 mode,
a mid-session resize and demotion, and both colour matrices.
Gates: container clippy -D warnings over four crates, 236 tests, workspace
check. pf-vkdecode and pf-bitstream are byte-for-byte untouched, so the
hardware rungs' 250/250 stands.
The libva AV1 layouts, the AuPlan conversion and the Linux rung's AV1 arm,
completing AV1 across all three hardware backends. Pin-only.
Layouts measured, not transcribed: the committed probe grew the AV1
structures and every size and offset it printed against libva 2.23.0 is a
compile-time assertion. Three that a hand-count gets wrong — the picture
buffer is align 8 because anchor_frames_list is a pointer, inserting seven
bytes of padding; seg_info and film_grain_info carry their own padding tails
inside the parent; and THREE of AV1's six bit-field unions are narrower than
a word (one uint8_t, two uint16_t), so a u32 packer over any of them writes
through its neighbour.
This is the fifth way this program has had to spell "which pictures does this
frame use", and it is unlike the other four: ref_frame_map is indexed by SLOT
and holds actual VASurfaceIDs rather than indices into anything, ref_frame_idx
is indexed by NAME and holds slots taken from the header — not from the
plan's refs, where a lost reference leaves a hole and a hole is not a slot —
global motion is picture-level, and there is no per-reference size field at
all. Established from va_dec_av1.h and libavcodec's vaapi_av1.c, and stated
in the module docs so the next reader does not re-derive it.
Review verified the whole happy path — every layout assertion re-measured,
every packer width and bit position, the reference convention, the
num_elements buffer shape — and found both defects on FAILURE paths, neither
reachable on the vendored vector.
A conversion refusal permanently desynced the ledger. The mutation block sat
after the tile walk, so any tile-shape refusal left the planner holding a
picture with no ledger slot — and the resulting UnresolvedReference fires
before that block too, so it never repaired. Every later access unit
hard-errored until a shown key frame: one lost packet costing a GOP. The
arm's own doc already warned that skipping conversion would desynchronise the
slot map; the refusal door did exactly what the skip door was written to
avoid. The block is hoisted, and a tile-shape refusal on an already-damaged
plan is now concealed rather than refused.
Fixing that exposed a sharper edge: the conversion can release a slot and
reassign it to the refused picture in one call, so the binding would still
hold the PREVIOUS picture's surface — a wrong reference rather than a missing
one, which nothing downstream could notice. The caller now clears the binding
unconditionally on the refusal path.
And a damaged frame's surface was never written yet was bound as a reference
and left in pending, so a later clean show_existing_frame would claim it with
damaged = false and ship uninitialised GPU memory to the presenter — on
several drivers another client's framebuffer. The justification quoted half
of va_dec_av1.h; its next sentence gives the remedy, which is to point the
problematic index at an alternative buffer. Damaged frames now submit as they
do on the other two arms, with live surfaces substituted for invalid entries
and reported as a bitmask — preferring a reference that really decoded over
the decode target, and keeping libavcodec's deliberate all-invalid map on a
shown key frame.
Film grain is refused rather than decoded wrong: libva wants two surfaces,
one ungrained for prediction and one grained for output, and libavcodec
allocates a second frame for exactly that. The gate now sits after the
mutation block so a grained frame costs itself rather than the GOP, and stays
per-AU rather than per-sequence because a stream that merely DECLARES the tool
decodes here perfectly.
⚠ Residual, flagged not fixed: a picture decoded from substituted references
can still be shown by a later show_existing_frame. It is decoded memory now
rather than uninitialised, and it is what the H.264/H.265 arms do, but
tracking "this was concealed" through to display needs new session state.
Gates: macOS fmt/clippy/125 tests/cargo-doc, container clippy -D warnings over
seven crates and 548 tests, workspace check. pf-bitstream's diff is
comment-only — verified — so the Vulkan rung's 250/250 stands untouched.
Nothing here has decoded a frame: no VAAPI hardware is reachable.
The AV1 arm of the native D3D11VA rung, parity-required because today's
FFmpeg d3d11va rung already decodes AV1 Profile 0 and the excision must not
silently drop it. Pin-only, as that rung is today.
decode() walks the temporal unit frame by frame; submit() splits into
decode_into and present, because AV1 decodes frames that are never shown. The
proven H.264/H.265 body is byte-for-byte unchanged — review diffed it against
HEAD mechanically and found only a rename plus one refusal arm — and the
VideoProcessorBlt hand-off is untouched. That mattered more than anything
else here: those two codecs are hardware-proven, .173 is powered off, and no
gate that runs could have caught a regression in them.
Every descriptor value comes from libavcodec's dxva2_av1.c read verbatim, not
from symmetry with the other codecs: three buffers and no qmatrix (AV1
transmits none), NumMBsInBuffer zero on all three, ConfigBitstreamRaw 1,
surface alignment 128, pool +8, and the session sized from the SEQUENCE
header's max frame size — sizing from the frame would rebuild the decoder and
drop every reference the first time a stream legally resized downward.
Two places where following the H.264/HEVC pattern would have been wrong.
libav pads the bitstream buffer and grows only its descriptor's DataSize,
never a tile's, because a tile's size is exact — charging padding to the last
record is corruption, not filler. And the committed tile records were one per
tile GROUP spanning the whole OBU, header and frame header included, where
libav emits one per TILE addressing the payload past its tile_size_minus_1;
the vendored vector is single-tile, so the old tests passed either way.
Review then found four more defects in the already-committed conversion, each
confirmed against libavcodec AND Chromium's D3D11 AV1 accelerator:
Tile widths and heights were the coded minus-1 where the field is a
superblock COUNT — every tile declared one superblock short, on every frame,
with a comment asserting the opposite of the truth.
StatusReportFeedbackNumber must be zero for AV1. Both reference
implementations disable it specifically for this codec — libav's note reads
"breaks decoding on some drivers (tested on NVIDIA 457.09)", Chromium's "it
crashes :|" — while both set it for H.264 and HEVC, which is why this rung's
proven codecs never showed it. It would likely have presented as a hang or a
rejected submission rather than bad pixels, sending the next session after
the tile records instead.
frame_refs[].Index is an index INTO RefFrameMapTextureIndex, not a surface
index; the neighbouring line already filled that map correctly. Measured:
1636 reference entries on the vendored vector where the two differ.
qm_y/u/v need the 0xFF "no matrix" sentinel — 0 is a valid matrix index, and
274 of 274 frames transmit no quantiser matrix, so every one was being
dequantized against matrix 0.
Also closed: the slot leak the Vulkan rung had already found and documented
(a frame refreshing no slot is never reported removed, so nine of them
exhaust the ledger); a tile-grid check that could not fire, replaced with
libav's own cols*rows guard; per-reference sizes now taken from the
reference's own header via RefState rather than the current frame's; and the
render size clamped against the decoded picture in both rungs, since AV1
permits a render size larger than the frame.
The parity leg was rewired through the real decode path — it previously
called the internals directly, so its hidden-frame assertion described the
harness's own counter rather than production withholding anything.
Gates: macOS fmt/clippy/383 tests, container clippy -D warnings over four
crates and 499 tests, and on Windows .133 (.173 is powered off) clean checks
plus 97 pf-dxvadec tests. All 8 Vulkan gpu_parity legs re-verified bit-exact
on the RTX 5070 Ti after the shared-code change.
No AV1 frame has been decoded through this rung anywhere: it needs .173 back.
The same use-after-free the AV1 rung was just fixed for, closed in the two
rungs that ship. session.rs and session_h265.rs handed their Std parameter
sets to vkCreateVideoSessionParametersKHR and dropped the backings when the
call returned; NVIDIA 610.57.04 was measured retaining such a pointer to
decode-record time, which is what made AV1 diverge on 250 of 250 frames.
Nothing was known to be broken here — both rungs are bit-exact on four
drivers — but that was luck rather than correctness: the freed blocks happen
to still hold the right bytes in that window. The native Vulkan rung sits in
the auto ladder above FFmpeg-Vulkan on shipping clients, so this was live
code, and its failure mode is silent wrong pixels rather than a crash.
StoredParams and StoredParamsH265 hold the parameters object together with
every wrapper it points at, so an object whose backing is gone cannot be
built. create_parameters_object takes the wrappers by value; the Add arm
adopts them only after a successful update, so a failed update drops what it
never stored; the Recreate arm replaces, destroys the old object, then drops
its backings, written explicitly so the ordering survives later edits. The
Add-vs-Recreate decision table and the VPS ledger are untouched — only
ownership moved.
params.rs still carried the refuted claim as a type-level contract, that
Vulkan "copies all parameter data before returning" and keeping the wrapper
alive across the call "is the whole obligation". Corrected to the measured
truth.
The tests are what stop this returning, and each was verified by sabotage:
inlining the H.264 PPS box fails at pps pScalingLists, inlining the H.265 SPS
DPB box fails at sps pDecPicBufMgr, and making either adopt drop instead of
store fails both session tests. Two lessons are recorded in them. Pointer
equality cannot be the assertion, because the Std struct carries pointers by
value and a stale one compares equal — the read-back is the discriminator, so
the tests clobber the dead stack first to make a dangling read deterministic
rather than lucky. And the first H.265 draft read six of eight pointers and
let the sabotage through, so it now reads every one with a labelled assert.
⚠ One site of this class remains, deliberately: the VkVideoProfileInfoKHR
chains, where wire()'s borrow dies with its enclosing block while the object
created from it lives on — three session creates, an image, a buffer, and a
query pool built from a raw pointer into a stack chain. It spans six modules
and all three codecs, and a profile is enums a driver resolves at create time
with no per-frame deref, so the risk is materially lower. It wants its own
pass with its own hardware verification.
Gates: macOS fmt/clippy/196 tests, container clippy -D warnings, pf-vkdecode
182/182 and pf-client-core 140/140. On the RTX 5070 Ti, all 8 gpu_parity legs
re-verified green after the change — H.264, H.265, Main 10 and AV1 all still
bit-identical to libavcodec.
250/250 frames bit-identical to libavcodec on NVIDIA 610.57.04, and all four
other parity legs (H.264, H.265, Main 10, both four-byte-prefix twins) still
green.
session_av1 built the sequence header, handed pStdSequenceHeader to
vkCreateVideoSessionParametersKHR, and dropped the backing the instant the
call returned — on the documented assumption that Vulkan copies parameter
data before returning. NVIDIA does not. It keeps the pointer and dereferences
pColorConfig when a decode is RECORDED. The freed block became our own next
allocation, whose bytes read back as mono_chrome = 1, and a monochrome frame
skips exactly loop_filter_level[2..3] (AV1 7.14).
That is the whole fingerprint two earlier rounds chased: luma bit-exact,
chroma off by small amounts, and rewriting the chroma levels in the bitstream
changing nothing — the driver read them correctly and then discarded them,
because it believed the stream had no chroma. StoredParamsAv1 now holds the
parameters object and its Std backing in one value, so an object whose
backing is gone is unrepresentable.
The road there is worth recording, because two well-evidenced conclusions
were wrong before this one was right. A software oracle reproduced the
divergence exactly by disabling chroma deblocking, and a GPU probe showed
chroma levels [8,12] and [63,63] producing byte-identical output — which
looked conclusive and was not. libavcodec's own Vulkan AV1 hwaccel is
bit-exact on this same driver, which proved the hardware fine and the defect
ours. ffmpeg never hits it: with VK_KHR_video_maintenance2 it uses inline
session parameters and never creates a parameters object at all.
The proof is direct rather than inferred: a throwaway Vulkan capture layer
dumped both submissions and every byte of our AV1 picture info already
matched libavcodec's, including the loop filter block; only the session
parameters layer differed. Watching the block's address showed correct bytes
at create and our next allocation at decode.
Ruled out on hardware, so nobody re-tests them: filmGrainSupport,
maxCodedExtent, maxDpbSlots/maxActiveReferences, VkVideoDecodeUsageInfoKHR,
the tile-start sentinel, the setup slot's SavedOrderHints, a NULL
pTimingInfo, and heap luck.
Two earlier fixes are confirmed against libavcodec's captured wire bytes and
kept: CDEF secondary strengths carry the coded value rather than the spec's
in-place fixup, and LoopRestorationSize is log2-based. The refuted
driver-ignores-chroma-levels claim is corrected everywhere it was written
down, and that probe test now passes and points at the lifetime of everything
a submission points at before blaming a vendor.
⚠ Adjacent and NOT fixed: session.rs and session_h265.rs drop their Std
backings the same way, and those sets carry embedded pointers too. Both are
measured bit-exact on four drivers, so nothing is known to be wrong — but the
contract now rests on a driver behaviour measured FALSE for AV1 on a shipping
driver. The SAFETY comments asserting it have been corrected; the structure
is deliberately untouched pending its own pass.
Gates: macOS fmt/clippy/336 tests, container clippy -D warnings, all green;
8/8 gpu_parity and 3/3 gpu_smoke legs verified on the RTX 5070 Ti.
WP0's real deliverable, and the hole every previous measurement in this
programme had. All prior timings ran against driver-zeroed buffers, so rate
control had nothing to code (~300 B/AU against an 833 KB quota) and only the
PIXEL-proportional half of the encode cost was ever exercised -- while the 4K60
HDR field report was a BITS/FRAME problem at 6.8 Mbit/frame.
Adds `pf_zerocopy::cuda::write_plane_from_host`, the exact mirror of the existing
read_plane_to_host. No new loader entry was needed: cuMemcpy2DAsync_v2 was
already in the table and CUDA_MEMCPY2D just needed the reverse memory types.
Linux-only by construction (pf-zerocopy's `imp` is cfg'd to linux).
⚠ Two harness mistakes found and fixed by looking at bytes/AU rather than
trusting the knob:
- Pure per-pixel noise is INCOMPRESSIBLE, so a low bitrate target does not
produce low bits/frame -- it OVERSHOOTS. At a nominal 50 Mbps the encoder
emitted 719 KB/AU against a 104 KB quota, and the three lowest rows of the
first sweep all sat at the same ~5.7 Mbit/frame. Sweeping nominal bitrate
measures nothing.
- So the sweep moves CONTENT DETAIL (block size) instead, and the x-axis is the
bits/frame the encoder ACTUALLY produced, never the one requested.
4K60 HEVC 8-bit, real content, single-engine vs forced-2:
bits/frame Ada 4090 Blackwell 5070 Ti
0.2-0.3 Mb 4567 -> 2381 1.92x 5549 -> 3552 1.56x
~1.1-1.2 Mb 5060 -> 2626 1.93x 5867 -> 4082 1.44x
~3.3 Mb 8478 -> 4455 1.90x 9286 -> 5862 1.58x
~9.6 Mb 16237 -> 8114 2.00x 16435 -> 9275 1.77x
RESULTS. (1) Encode time scales strongly with bits/frame -- 4.6 ms to 16.2 ms
across the range on Ada -- confirming the hypothesis' core claim. (2) There is NO
CROSSOVER: split wins at every point on both architectures (Ada ~1.9-2.0x and
notably flat, Blackwell 1.44-1.77x). So the arbitration's encode-side answer is
essentially always "split", which makes the sub-frame handicap the only decision
that actually matters -- exactly the part already built and unit-pinned.
(3) It corroborates the field capture: at ~6.8 Mbit/frame these curves put
single-engine 4K60 around 10-13 ms, and the field report was 10.3 ms on a 4090.
That reads as real ASIC time, not the retrieve-queue inflation it might have been.
⚠ Caveat the data itself shows: cost is NOT monotonic in bits/frame alone. The
1px row lands at the HIGHEST bits/frame yet encodes FASTER than the 4px row on
both boxes (Ada 10148 vs 16237 us) -- pure noise defeats motion estimation, which
gives up early, where semi-structured content makes it search hard. Content
structure is a real term, so "bits/frame" is a good axis but not a complete cost
model.
Verified .21: clippy -D warnings clean (pf-encode + pf-zerocopy), 64 unit tests,
25/25 NVENC on-hardware. Curves run on both Ada and Blackwell. fmt clean.
WP1.3, and the measurement that justifies it. `resolve_split_mode`'s 10-bit rule
sat ABOVE the pixel-rate arm and took no codec, so it (D1) vetoed 10-bit 4K120 --
the very case the pixel-rate arm exists for -- and (D2) applied an HEVC-Main10-on-
Ada result to AV1 10-bit, which has no such measurement. Both fixed: the
pixel-rate arm now comes first, and what remains is codec-scoped to HEVC and only
applies BELOW that bar, where a second engine buys nothing anyway.
The rule rested on one datapoint: 5120x1440@240 Main10 on Ada, forced-2 7.6 ms
vs 2.8 ms single-engine -- split 2.7x SLOWER. Dropping the short circuit flips
that exact configuration's behaviour, so it was re-measured on a 4090 (AD102,
driver 610.43.03), 400 Mbps, sub-frame pinned off, via a new mode-parameterizable
Main10 A/B test (PF_AB_MODE=WxHxFPS reproduces the original operating point).
Ada 4090 single forced-2 ratio
3840x2160@60 4483 us 2178 us 2.06x split WINS
5120x1440@240 3689 us 2813 us 1.31x split WINS <- the veto's origin
3840x2160@120 4148 us 2189 us 1.89x split WINS
Blackwell 5070 Ti
3840x2160@60 4216 us 2477 us 1.70x split WINS
5120x1440@240 4651 us 3894 us 1.19x split WINS
Split wins for Main10 at every mode on BOTH architectures, including the config
the veto came from. The original number does not reproduce.
⚠ Caveats, unchanged from the rest of this work: content is trivial (297-300 B/AU
against an 833 KB CBR quota -- zeroed VRAM), so this is the pixel-proportional
term and the bits/frame regime is still unmeasured; debug build; and the driver
differs from whenever the original was taken.
Also validated on Ada in the same session -- the whole spike set reproduces on a
SECOND architecture and an OLDER driver (610.43.03 vs 610.57.04): S1a in-place
split switch accepted with zero IDRs both directions; S1b takes effect
(|C-B|=12 vs |C-A|=1921, the cleanest run yet); S1c pair flip passes; D5 confirmed
(AUTO+sub-frame 4424 vs DISABLE 4409, 15 us apart -- and AUTO without sub-frame
2310 ~= TWO_FORCED 2314, so the arm stays); engines=2 with THREE_FORCED correctly
clamped to mode 2; arbitration converged with exactly 1 keyframe.
Verified: .21 clippy -D warnings clean + 64 unit tests; .133 Windows clippy
-D warnings clean (the resolver signature grew a `codec` param, so both backends
moved); Ada + Blackwell on-hardware as above. fmt clean.
The named next step after WP3's first increment. That increment deliberately
REFUSED to arbitrate HEVC-with-sub-frame -- the fleet default, and the reported
field case -- because engaging split there gives up sub-frame readback, whose
whole value is that the send overlaps the encode. An encoder measuring only
encode time would see split as ~2x faster, take it, and make end-to-end latency
worse while reporting a win. This supplies the missing number.
The real comparison is encode_1eng + send_of_last_slice against
encode_2eng + send_of_whole_AU, so the challenger owes roughly
spread x (slices-1)/slices. Split across the two sides that can each see half:
- Host: new `Encoder::set_send_spread_us` (defaulted, forwarded by
TrackedEncoder -- same trap class as set_wire_chunking, and unforwarded it
would fail SILENTLY IN THE SAFE DIRECTION, which is the hardest kind to
notice). The send thread is the only place a paced send is observed and the
encode loop the only place the encoder can be touched, so it goes over an
AtomicU32 like encoder_ceiling_kbps, EWMA-smoothed 3:1 per completed AU: one
content spike must not flip a verdict that then gets cached.
- Encoder: turns the raw spread into the handicap, because only it knows
`slices`. SplitArbiter::with_handicap charges it to the challenger before the
comparison. A unit test runs identical encode numbers with a cheap and an
expensive send and asserts the verdict REVERSES -- with an expensive send the
arm that looks twice as fast is a loss end to end, and the incumbent must
hold. That is precisely the regression an encode-only arbiter ships.
Gate now opens for HEVC+sub-frame only when a spread has actually been reported
(and slices >= 2); with no hint it still refuses, so behaviour is unchanged until
the host feeds it.
Two mechanics this needed:
- apply_split_mode became a PAIR flip (split + sub-frame), routed through
resolve_split_subframe and restoring from `subframe_opened_with` so a session
that never had sub-frame can never gain it. It also recomputes
`subframe_chunks`, which reconfigure_bitrate does NOT -- spike S1c's finding;
leave it stale and supports_chunked_poll keeps saying yes while numSlices never
advances, so poll_chunk busy-polls its whole budget every AU.
- The arbiter is now fed from BOTH completion points. A sub-frame session
finishes through poll_chunk, so the incumbent arm of an HEVC experiment would
otherwise never deliver a sample -- only the challenger, with sub-frame
dropped, comes through poll.
Verified .21: clippy -D warnings clean for pf-encode AND punktfunk-host with
nvenc, 63 unit tests (1 new), 23/23 NVENC on-hardware green. Verified .133:
Windows clippy -D warnings clean, zero dead_code. fmt clean.
The fix S1 unlocked. Rather than predict the right split mode at open — which
cannot work, because the decision depends on bits/frame and an Automatic client's
steady-state bitrate is unknown at open (ABR climbs in place afterwards) — the
encoder now measures both arms on the live session and keeps the winner. S1
proved nvEncReconfigureEncoder takes a changed splitEncodeMode with
resetEncoder=0, emits no IDR, and actually applies it, so the experiment is
invisible on the wire.
Deliberately measures instead of modelling: hard-coded per-arch constants are
exactly how the rule this replaces went wrong (one 5120x1440@240 Ada datapoint
generalised into a fleet-wide 10-bit veto). A measurement tracks driver updates
for free.
`SplitArbiter` (pure state machine, unit-tested without a GPU): measure incumbent
-> switch -> SETTLE -> measure challenger -> keep the winner, else switch back.
Verdicts cache per (gpu, codec, mode, depth, chroma) so later sessions open
straight into the winning arm; the key is CeilingKey minus split_mode, since the
split mode is the thing being decided.
⚠ SETTLE_FRAMES=16 is load-bearing, not padding: split-encode does not reach
steady state on the first frame (a FRESH TWO_FORCED session measured early-half
3280us vs late-half 1996), so judging an arm right after switching reads the
transient — intermittently, which would then be cached. A unit test feeds exactly
that transient and asserts the arbiter still sees the steady state.
Safety gates, all correctness conditions rather than preferences: opt-in
(PUNKTFUNK_NVENC_SPLIT_ARBITRATE=1) while it earns trust; an operator
PUNKTFUNK_SPLIT_ENCODE pin always wins; skip if a verdict is already cached; sync
depth-1 only (async_rt.is_none(), same gate chunked poll uses — under pipelined
retrieve the submit->AU span includes queue depth and the comparison is noise);
needs >=2 engines; never H.264.
⚠ And the one that bounds this increment: NO SUB-FRAME TRADE. For HEVC, forcing
split gives up sub-frame readback, which costs send/encode overlap the ENCODER
CANNOT SEE — it measures encode time only, so it would reliably prefer split and
silently make end-to-end latency worse. So arbitration runs only where nothing is
traded: sub-frame already off, or AV1 (both features legal). Pricing that trade
needs the host's send cost and is the next work package.
Challenger choice tests the question worth asking — anything not already the
widest forced split is challenged BY the widest ("are we leaving engines idle?").
The naive "challenge whatever we are not" spent the experiment re-proving that
splitting beats not-splitting, while parking the session on the slow arm to do
it, because 4K60 sits on the fallthrough AUTO.
⚠ Every new nvenc_core item is linux-gated: the arbiter is wired into the Linux
backend only for now and nvenc_core compiles on Windows too. Caught by the .133
check, not by reasoning — the first cut failed Windows clippy with 12 dead_code
errors, the exact item-level trap this file already carries a scar from.
Verified .21: clippy --features nvenc --all-targets -D warnings clean, 62 unit
tests (4 new arbiter tests), 23/23 NVENC on-hardware green including a new
end-to-end convergence test asserting ZERO extra IDRs and a cached verdict.
Verified .133: Windows clippy -D warnings clean, zero dead_code. fmt clean.
"unom" alone is not the legal entity name. Updated the copyright line in
all nine first-party license files -- the root MIT/Apache pair and the
hand-maintained copies under clients/apple, clients/decky,
packaging/windows/drivers and packaging/windows/pf-vkhdr-layer (there is
no script that syncs these, so each is edited directly).
The Linux and Windows clients' About screens pick this up automatically:
both `include_str!` the root LICENSE-MIT / LICENSE-APACHE at compile time.
Deliberately untouched:
- Third-party license texts (Geist OFL, FFmpeg, VB-CABLE, the vendored
pyrowave/Granite tree, KDE protocol XMLs, the os-icon licenses) -- those
are other parties' copyrights.
- Publisher/author/maintainer metadata, which is identity rather than
license text and is reported separately for a decision.
The docs site's page title read "punktfunk docs". Fixed that and swept the
rest of the tree for the same defect, capitalizing the brand wherever it is
shown to a human and leaving it lowercase where it is a technical identifier
(CLI/package names, `punktfunk://` scheme, PnP enumerator, TLS SNI, logcat
tag, config paths, CMS tenant id).
User-visible fixes:
- docs-site: page title -> "Punktfunk Docs"; API reference title, meta
description and the branded bar's aria-label; the BrandMark/Wordmark SVG
accessible names (the web console already had these capitalized -- the
docs site had drifted from it).
- Android: six strings of live UI copy -- the local-network permission
dialog (x2), the connect-screen error banner, and the no-controller
explainer.
- Apple: the "No Hosts" empty-state text and the fallback display name for
a host that advertises no instance name.
- Windows client: the `--discover` progress line.
- KWin fake-input: the application name passed to `authenticate()` (the
grant is cached per-exe, so the string is display-only).
- THIRD-PARTY-NOTICES: fixed in both generators (about.hbs and
gen-third-party-notices.py) and applied to the three checked-in outputs
so they match what a regeneration now produces.
Every changed line differs from the original only by letter case, so line
lengths are unchanged and no formatter width rule is affected.
`cargo fmt --all --check` passes.
A parity and smoke harness for AV1, mirroring the H.264 and H.265 legs that
proved those rungs bit-identical to libavcodec on four drivers before either
ran on glass. This was the milestone's largest test gap: the adversarial
review found four blocking defects in the AV1 conversion — flags unset on
274 frames of 274, a units error in LoopRestorationSize, per-reference info
describing the wrong picture, film-grain fields left zero — and every one of
them would have shown on frame 1 of a parity run, while clippy and 164 green
unit tests said nothing at all.
The golden is 250 per-frame SHA-256s in DISPLAY order, not 274. The vector
carries 274 coded frames in 250 temporal units; the 24 extras are hidden
ALTREFs, decoded and referenced but never shown, and the rung delivers what
dpb.outputs names. The count is re-derived from the planner rather than
assumed.
Cross-checked between ffmpeg 8.1.1 on macOS arm64 and 8.0.1 on Linux x86_64,
whose raw outputs are byte-identical — and then against a third party neither
build knows about: the vendored vector ships upstream's own per-frame MD5s,
and re-running those reproduces all 250. The golden agrees with a decode
nobody in this program performed. I reproduced both independently before
committing.
8-bit NV12, traced from the sequence header rather than presumed
(seq_profile 0, high_bitdepth 0, mono_chrome 0), so the P010 scar does not
apply here — and the header says which check to make if a Main 10 golden is
ever added. film_grain_params_present is 0, which is load-bearing: grain
synthesis is part of the Vulkan decode profile, so this golden is only
comparable against a grain-less profile key.
Anti-vacuity is the point of the exercise, so it is structural. The golden
guard asserts the exact count, that every line is a bare digest, and that all
entries are DISTINCT — 250 copies of one digest would let a decoder frozen on
a single frame pass parity. The parity body asserts the golden set and the
access-unit count before it touches hardware, so an IVF reader returning
nothing cannot become "0 frames compared, pass". The agent verified the
guards fire by mutating the golden three ways.
assert_bit_identical now names the FIRST divergent frame, which is what
localises a defect; that improves all six legs, not just AV1.
AV1 has no four-byte-start-code twin, deliberately: OBUs are
length-delimited, so there is no prefix for a driver to mis-skip. Documented
where a reader would otherwise see an omission.
Nothing here has run on a GPU. The harness exists precisely so the four
review defects can be answered by measurement instead of argument.
Two separate things had to be wrong for this, and both were.
The frame's own policy locked it out. Plugin UIs moved to their own origin so a
plugin cannot act as the logged-in operator, and the plugin origin names the
console as the only page allowed to frame it. It built that name from the
incoming request — but Nitro hands the app a synthetic request with no TLS
socket, so an HTTPS console reads back as `http:`. The header said
`frame-ancestors http://host:47992` while the operator was on
`https://host:47992`, and the browser refused the frame outright
(ERR_BLOCKED_BY_RESPONSE). Nothing on screen said so; the reason was only in
devtools. The scheme now comes from the listener's own TLS state, stamped at
bind time, with x-forwarded-proto winning when something in front terminated
TLS for us — the one case where the browser's scheme is not ours.
And the port was shut. 47993 was added to the firewall definitions, but an
already-open firewall does not pick a new port up: ufw expands an app profile
into rules when you allow it and keeps those, so editing the profile on upgrade
changes nothing, and firewalld needs a reload. Every upgraded Linux host kept a
47992-only rule, silently. The packages now notice on upgrade and print the one
command that fixes it, without touching the running firewall. The NixOS module
and the container image never learned the port at all; both now open it.
Also: the console no longer mounts the frame while it is still checking whether
that origin is reachable. A firewalled port drops rather than refuses, so the
check hangs for the browser's whole connect timeout, and mounting meanwhile is
exactly the empty panel with no explanation. The card that follows now names
both causes it can be — an untrusted certificate for that port, or a closed one
— because from a browser the two are indistinguishable.
The rule is now a pure function with tests, since its failure mode is a
well-formed header that only a browser rejects.
Verified on glass against home-worker-5 (.21) and its ROM Manager plugin: the
frame was refused before, renders the plugin's UI after.
Last change's docs concluded "AUTO never splits, retire the arm" from the
sub-frame-ON measurement alone. Measured the missing leg before implementing it,
and the conclusion was wrong.
On .21 at 4K, plain AUTO (env unset, the resolver's fallthrough):
sub-frame ON -> 5023/5157 us/frame ~= DISABLE 4979/5000 (does NOT split)
sub-frame OFF -> 2401/2352 us/frame ~= TWO_FORCED 2319/2378 (DOES split)
So AUTO is CONDITIONAL, not dead. Retiring it would have silently cost every
sub-frame-off session its second engine -- a regression introduced while
"cleaning up" an arm that looked inert. Split and sub-frame are mutually
unsupported for HEVC, so the driver resolves AUTO to no-split only in that
combination.
Fix is disclosure, not removal:
- resolve_split_subframe debug-logs the inert HEVC + AUTO + sub-frame case, which
is the fleet default shape: "split_mode=AUTO" has meant "no split" for every
default session and nothing said so. Deliberately NOT rewritten to DISABLE --
the mode we pass is what the driver was actually given, and the ceiling-cache
key must keep describing that.
- New unit test `auto_survives_the_arbitration_in_both_subframe_states` pins the
contract so the arm cannot be simplified away later.
- The resolver doc now records both measured legs instead of "AUTO is dead".
Also in this change:
- WP1.6: `resolve_subframe`'s doc said "Windows passes `false`". Stale since the
2026-07-31 .173 A/B flipped Windows to caps-gated default-on. It mattered:
it made the AUTO-plus-sub-frame dead combination look Linux-only when it is
fleet-wide.
- Windows session-ready log parity: split_mode + engines + subframe. The Windows
line had no split_mode at all, so a Windows field report could not answer "did
this session actually split?" -- the question that started this whole thread.
Verified: fmt clean; .21 clippy -p pf-encode --features nvenc --all-targets
-D warnings clean, 58 unit tests (1 new), 22/22 NVENC on-hardware tests green;
.133 Windows clippy --features nvenc --all-targets -D warnings clean (15m cold,
zero errors or warnings) -- the Windows backend is cfg'd out on both macOS and
the Linux box, so that leg needed a real Windows host.
The third codec arm in video_vk_native, AV1 admitted to native_codec and to
native_vulkan_gate by pin only. It stays out of `auto` on the same rule M5's
D3D11VA rung follows: `auto` admission is earned with hardware evidence, and
this has decoded nothing on a device.
is_integrity_warning_av1 did not exist, so the client could not have
concealed AV1 damage at all. Added, exhaustive, no wildcard: all three AV1
warnings really are damage, because AV1 has no spec-legal-but-noisy signal
to mis-classify — no reorder envelope to announce, no MMCO to rebase — and
the exhaustive match is what stops a future variant defaulting to clean.
The blocking defect review found was two safety mechanisms cancelling each
other. After a failure the decoder skipped to the next key frame answering
Ok(None), and because AV1's planner has no flush its store kept planning
cleanly, so those AUs carried no warnings and the client read them as proof
the rung works — clearing the demotion streak and resetting its clock on
every one. The streak could then never reach the threshold, which made the
never-delivered fall-through to FFmpeg-Vulkan unreachable, which is the
documented backstop for exactly three things: a level above maxLevelIdc, a
sequence header disagreeing with the Welcome, and film grain. Film grain is
the probe's own admitted assumption, so a grain stream would have frozen the
screen for the session while DecodeHealth reported run 0 — recovered.
AV1 now answers the wait with an error, as H.264 and H.265 already do
through AwaitingIdr, so all three codecs are indistinguishable to the
demotion machinery. That matters more than the extra precision of a third
state: only the H.26x paths have hardware evidence, and they are proven WITH
that behaviour.
The obvious form of that fix would have wedged the decoder. A key frame can
sit behind a skipped frame inside the same temporal unit — the vendored
vector has 24 two-frame units — so erroring out of the per-plan loop would
never reach it and the wait would never end. Skips are therefore counted per
frame and the error raised only when the whole unit was skipped, with the
metadata-only unit staying a clean Ok(None).
Also closed: a refused temporal unit left an already-decoded frame in the
ready queue, which shipped on the next AU as a clean success — putting a
picture from a refused AU on screen, clearing the streak again, and latching
delivered so the fall-through was disabled for good. The error arm now
drains and releases unshown.
MAX_DELIVERABLE is derived rather than picked: HOLD_HEADROOM minus the
pipeline's own hold, pinned to pf-vkdecode's constant so a hardcoded depth
fails the build. At the previous 8 the queue plus the presenter's 4-7 stood
against a headroom of 8, so it capped memory without preventing the
exhaustion it named, and a frame waiting 8 AUs burned 16 of the 17 query
slots — where a re-armed slot reads as Failed and becomes a fabricated
driver-corruption verdict in the very counter the Ally X signal lives in.
The trim now runs after this AU's frame is taken, or at the derived depth it
would drop a two-output unit's first frame and invert display order inside
one AU.
Its justification was also wrong: the claim that a temporal unit may carry a
show_existing_frame alongside a shown frame is disproved by this repo's own
golden — 250 units, 250 shown, zero show_existing. The bound is kept as
defence in depth against a non-conformant or multi-operating-point stream,
and now says so.
Gates: macOS fmt/clippy/392 tests, container clippy -D warnings over six
crates, 851 tests, workspace check. No hardware: the rung is pin-only and
has still never decoded a frame on a device.
WP1.1 plus the engine-count fix. `resolve_split_mode` forced TWO_FORCED at high
pixel rate regardless of hardware, so a 3-NVENC part (GB202, AD102 workstation)
left a third of its encode silicon idle, and a 1-NVENC part paid a wasted session
open to discover it could not split.
Probes NV_ENC_CAPS_NUM_ENCODER_ENGINES in both direct-SDK backends' query_caps
(the cap is `= 49` in both linux_sys and windows_sys of the vendored SDK 0.4.0 --
the caps enum is cfg-selected per-OS, so that was checked) and latches it on a
backend field. NOT on EncoderCaps: nine backends construct that struct as
exhaustive literals, so a new field would be a 9-site change of which 7 are
unrelated codecs passing a meaningless value, and the only consumer is the
resolver.
New `max_forced_split_mode(engines)`: 1 -> DISABLE, 2 -> TWO, 3 -> THREE, and
>3 -> AUTO_FORCED, because NV_ENC_SPLIT_ENCODE_MODE cannot NAME more than three
(NVENCAPI 12.1; values 4..14 are unallocated, so a future API may extend it) and
AUTO_FORCED = "split, driver picks how many" is measurably a real split (2.01x vs
disabled on .21). 0 = unprobed keeps the historical two-engine assumption.
⚠ WHY THE CLAMP EXISTS, measured on .21 (RTX 5070 Ti, 2 NVENC, 4K HEVC):
requesting THREE_FORCED was HONOURED -- session opened in mode 3 -- and ran at
2303 us/frame, identical to TWO_FORCED's 2308. The driver does not reject an
over-ask; it silently encodes narrower. So the rejection fallback cannot find the
ceiling and PUNKTFUNK_SPLIT_ENCODE=3 on a 2-engine card would have logged a
3-way split over a 2-way encode. Operator overrides are now clamped with a warn.
The ordering trap is covered by a test: on a >3-engine part hw_max is
AUTO_FORCED (1), which is not "narrower than" TWO_FORCED (2) despite comparing
smaller, so a naive min() would collapse a legitimate 3-way request to AUTO.
Also adds `engines` and `subframe` to the Linux session-ready log: split_mode
alone is ambiguous between "used both engines" and "left a third idle", and
since the driver honours an over-wide request the mode cannot be read without
the ceiling it was chosen from. This is the line a field report needs.
--- and a correction to S1b, in the same change ---
Re-running S1b afterwards flipped its verdict to "the driver appears to have
IGNORED the in-place split change", contradicting the isolated runs that produced
the |C-B|=34 figure already written into the design docs. Investigated rather
than re-rolled.
The switched leg was landing MIDWAY between the arms (~3600 us against A~5050,
B~2300) and the nearest-neighbour verdict flipped on noise. Cause: split-encode
does not reach steady state on the first frame -- a FRESH TWO_FORCED session
shows it too (early-half 3280 us vs late-half 1996 in one run), so it is split
warmup generally, not something specific to reconfiguring in place. A single
median over the whole window cannot see that. The test now reports early-half vs
late-half and gives a switched leg SETTLE=16 frames before its window opens,
every leg the same length. With that, 4/4 runs agree: the switched leg reaches
~2030 us against a fresh-split ~2000 and a single-engine ~4900.
⚠ S1b's CONCLUSION stands (the switch does take effect) but the evidence behind
the committed number did not reproduce; the docs are corrected rather than left
implying a cleaner result than the harness could support.
⚠⚠ This is a WP3 REQUIREMENT, not just a test fix: a live-session arbitration
that switches arms and immediately measures will misjudge the arm it just chose,
because the encoder needs ~16 frames to settle. The settle window has to be part
of the arbitration, and it is now a measured number rather than a guess.
Verified on .21: clippy --features nvenc --all-targets -D warnings clean,
57 unit tests (3 new), all 23 NVENC on-hardware tests green, fmt clean. The 3
failing on-hw tests in a full --ignored run are VAAPI (no AMD/Intel GPU on that
box -- their own ignore reason says so), pre-existing and unrelated.
S1c `nvenc_cuda_split_subframe_pair_reconfigure`: the leg S1a/S1b excluded. Both
pinned sub-frame OFF to isolate the split variable, but a real HEVC arbitration
cannot -- split and sub-frame are mutually unsupported there, so engaging split
means flipping enableSubFrameWrite in the same breath, a second init param and
the one the reconfigure path deliberately pins. RESULT on .21: the PAIR moves in
place, accepted, ZERO IDRs, both directions.
It also pins the invariant that makes this safe to build on: `subframe_chunks` is
latched ONLY in the init path (~line 1625) and is NOT recomputed by
reconfigure_bitrate, so a caller flipping sub-frame in place must clear it too or
supports_chunked_poll keeps reporting true and poll_chunk busy-polls its whole
budget every AU against a numSlices that never advances. The test performs the
correct sequence and asserts the state stays coherent, so WP3 has a worked
example rather than a warning.
`nvenc_cuda_auto_split_with_subframe`: the D5 confirm -- the one claim in the
design's defect list that was only ever inferred. The driver reports no "mode I
actually chose", so it is settled by timing, at 4K where the gap is ~2x.
RESULT: AUTO (env unset) + sub-frame 4904 us/frame, DISABLE + sub-frame 5062,
TWO_FORCED without sub-frame 3464. AUTO sits 158 us from DISABLE and 1440 from
TWO_FORCED ⇒ D5 CONFIRMED: plain AUTO does not split while sub-frame is on, so
the resolver's AUTO fallthrough reads as "let the driver decide" and means
"never split".
⚠ TRAP, hit on this test's first run and now documented in it: the env knob
CANNOT express plain AUTO. `0` is DISABLE and `1` is AUTO_FORCED, and
resolve_split_subframe counts AUTO_FORCED as forced, so passing `1` silently
disarms sub-frame and measures a different configuration entirely -- which
produced a spurious "D5 REFUTED". Plain AUTO is only reachable as the resolver's
fallthrough with the env unset. The leg now asserts sub-frame resolved TRUE, so
the test can no longer answer the wrong question quietly.
Verified on .21: clippy --features nvenc --all-targets -D warnings clean, all 4
spikes green, the normal 54-test suite unaffected, cargo fmt --all --check clean.
Android was the one app that could open a punktfunk:// link but never hand
one out, so every Android link had to be typed by hand — and the host's
stable record id, which is the part that keeps a link working after the box
changes address, isn't shown anywhere in the UI to type.
Both homes now offer Copy link: the touch grid's card overflow menu, and the
controller home's host options (Up on a tile). A pinned card copies its own
profile with it, matching Linux and Apple; a host card copies none and so
keeps honouring the host's binding, exactly like tapping it does.
The URL is the shared self-emitted form (DeepLinks.forHost), already covered
by the cross-language vector tests, so the three emitters stay in step.
Android 13+ draws its own clipboard confirmation and we add nothing on top of
it; below that we say so ourselves, as a toast in the console home, which
renders neither banner.
caps_av1 / session_av1 / decoder_av1, over the CPU half already committed,
sharing the picture pool, bitstream ring, op ring, DPB settling and frame
delivery with H.264 and H.265 rather than forking them. AV1 session
parameters carry exactly one sequence header — no PPS, no VPS — so the
parameters ledger is two-state: current, or recreate.
The GPU plumbing came through review clean. The damage was all in the
conversion committed two rounds ago, which nothing tested against a
reference, and none of it would have failed a gate: clippy was clean, the
tests were green, and the rung would have decoded its own conformance vector
wrong on essentially every frame on AMD, silently.
Four blocking defects, each measured on the vendored vector rather than
argued:
Nine StdVideoDecodeAV1PictureInfo flags were never set. Four change
reconstruction — allow_screen_content_tools on 274 frames of 274,
allow_warped_motion on 273, is_filter_switchable on 172, force_integer_mv on
1 — and RADV reads three of them directly. The block already set
allow_intrabc, which is only codeable when screen-content tools are on, so
it contradicted itself.
LoopRestorationSize sent the pixel size where the field is log2(size) - 5.
cros-codecs stores 64/128/256; RADV names its destination
log2_restoration_size_minus5 and reads 1/2/3. Nothing truncates, nothing
errors, and every frame with loop restoration reconstructs against a
nonsense unit size.
Per-reference Std info answered questions about the wrong picture: every
reference carried the CURRENT frame's type, and RefFrameSignBias was never
set at all. Sign bias is what tells a decoder a reference lies in the
future, and this vector is the hidden-ALTREF one, so all-zero meant every
reference was treated as past. Fixed at the source: pf-bitstream now records
a RefState when a picture is stored — its own frame type, sign-bias mask,
saved order hints — and carries it on the slot, so all three backends get
answers about the reference rather than about the frame reading it.
Film grain's six chroma-scaling fields were zero, which defeats the profile
machinery that exists to refuse devices unable to synthesise grain.
The reference-name compaction is fixed in the PLANNER, once. AuPlan::refs is
now name-indexed with holes preserved, so a lost reference can no longer
renumber every later AV1 reference name — a class that was live in both
conversions and armed for the VAAPI rung that does not exist yet. The DXVA
twin had a second name-versus-slot confusion: it read global motion by DPB
slot from an array the spec indexes by reference name, and slot 0's matrix
is all-zero rather than identity, so 273 references were given a zero warp.
Also closed: pTileOffsets/pTileSizes were sized to tileCount while RADV
reads AV1_MAX_NUM_TILES entries unconditionally — a 4-byte allocation read a
kilobyte deep — now fixed 256-entry arrays with zeroed tails. And the test
guarding the lost-reference refusal re-implemented the predicate inline, so
deleting the guard left it green; both now call one named function.
The bitstream layout now matches libavcodec: raw tile payloads only,
frameHeaderOffset 0. The review established the spec-literal layout was NOT
wrong — AV1 has no start-code scanning, so the 3-versus-4-byte and
slices-only scars do not transfer, and no driver in the fleet reads
frameHeaderOffset — but matching the validated reference deletes code,
uploads 5835 fewer bytes over the vector, and removes the untested-driver
tail.
Upstream, and the third of its kind: the vendored parser writes
ref_frame_sign_bias[i] in the same loop body where it writes
order_hints[LAST_FRAME + i], so its array is shifted one down and index 7 is
never written. Corrected in RefState::of with the shift documented, the
vendored tree untouched, and pinned by a test that recomputes the bias from
order_hints through the parser's own get_relative_dist.
Gates: macOS fmt/clippy/tests, container clippy -D warnings over six crates,
845 tests, workspace check. No hardware: nothing here has reached a driver.
Two on-hardware spikes answering the gate on the split-encode engagement
program (design/nvenc-split-encode-engagement-implementation-plan.md).
S1a `nvenc_cuda_split_reconfigure_in_place`: can splitEncodeMode change via
nvEncReconfigureEncoder with resetEncoder=0, without an IDR? Our "reconfigure
must present the SAME init params as the open" rule (windows/nvenc.rs:620) is
our own invariant and had never been tested against a driver. It reports rather
than asserts the verdict -- both outcomes are legitimate findings -- and only
asserts what would invalidate the measurement (session live, engines >= 2, the
arms actually differ). Sub-frame is pinned off so the driver can't reject for
the wrong reason (HEVC forced-split and sub-frame are mutually unsupported).
S1b `nvenc_cuda_split_reconfigure_takes_effect`: the other half -- a driver that
accepts the parameter and quietly ignores it looks identical to one that honours
it. Three legs at 4K (fresh DISABLE / fresh TWO_FORCED / DISABLE->TWO in place);
if C tracks B and not A, the switch is real.
RESULT on .21 (RTX 5070 Ti, GB203 Blackwell, driver 610.57.04):
NV_ENC_CAPS_NUM_ENCODER_ENGINES = 2
S1a: accepted, ZERO IDRs, both directions.
S1b: A fresh DISABLE 5054 us/frame, B fresh TWO_FORCED 2453,
C switched in place 2419 -- |C-B|=34 vs |C-A|=2635. It takes effect,
and split is a clean ~2x at 4K.
Two limits, both recorded in the test docs rather than the commit only. The
frames come out at 427 B/AU against an 833 KB CBR quota: the driver hands back
zeroed VRAM, so the rotated buffers are identical and rate control skip-codes
everything. So this measures the PIXEL-proportional half of the cost only --
the bits/frame regime the field case lives in is untested here, and the test
prints an explicit INCONCLUSIVE-on-content line when it detects that. And this
is Blackwell 8-bit; the Ada Main10 question is untouched.
Verified on .21: clippy -p pf-encode --features nvenc --all-targets -D warnings
clean, both spikes green, cargo fmt --all --check clean.
The standing open item M7 was meant to close. `decodable_codecs` answered the
AV1 bit from `ffmpeg::decoder::find(AV1)`, which says yes on every build that
links libdav1d — a software decoder. So the client told the host "send me
AV1" on machines that would then decode a 4K stream on the CPU, and codec
negotiation happens once at Welcome, so there is nothing to fall back to
afterwards. A promise the client cannot keep is worse than not making it.
`av1_hardware_decodable` answers from device facts only: the presenter's
Vulkan device listing DECODE_AV1 among its decode queue family's codec
operations, or — on Windows — the D3D11 import path, which is the same gate
the D3D11VA rung sits behind and that rung decodes AV1 Profile 0 today.
VAAPI is deliberately not consulted: asking libva costs opening a display,
and this is called too early and too often for that. The Vulkan bit covers
the Mesa devices where VAAPI AV1 exists in practice, and a machine with
VAAPI AV1 but no Vulkan AV1 loses the advertisement, not a working path.
The test pins what the gate must not accept: a device that decodes H.264 and
H.265 but lists no AV1 operation, and a device whose caps word claims AV1
while it has no decode queue at all.
Gates: macOS fmt/clippy, container clippy -D warnings over six crates, 805
tests, workspace check.
One AuPlan into DXVA_PicParams_AV1, over the layouts the SDK header measured.
AV1 on DXVA needs TWO reference arrays that mean different things at once,
and this program has now written down four spellings of the same question.
`frame_refs[7]` is indexed by reference NAME and each entry carries a
SURFACE index — where Vulkan's `referenceNameSlotIndices` carries a SLOT —
plus that reference's own global motion. `RefFrameMapTextureIndex[8]` is
indexed by SLOT and states the whole reference store, which is what
`RefFrameList` is for the other codecs and why a long-term reference no
frame names still has to appear in it.
The test asserts that difference is exercised rather than assumed: it fails
if the run never saw the store hold a picture the frame did not name, which
is precisely the distinction the Ally X class of bug lives in.
Three transpositions that would each have been silent:
Global motion is signalled per reference SLOT in the frame header and stored
per reference NAME in DXVA, so the conversion reads by one and writes by the
other. Carrying the Vulkan shape across would leave every warped reference
at identity.
CDEF strengths pack two fields to a byte, primary in the low six bits and
secondary in the top two, where the AV1 syntax keeps parallel arrays.
DXVA wants log2 of the loop-restoration unit size; the parser records the
size. And the superres denominator is the real one here — SUPERRES_NUM when
superres is off — where Vulkan's `coded_denom` is the denominator less nine.
Film grain rides only where the sequence enables it and the frame applies
it, its scaling points transposed into [value, scaling] pairs, and an
over-count refused rather than truncated: fewer points than the stream
declared is different grain, not less of it.
Gates: macOS fmt/clippy/349 tests, container clippy -D warnings over six
crates, 804 tests, workspace check.
The blocker on this rung was never the code — it was that `dxva.rs` is the
most safety-critical file in the backend and nothing in it is type-checked
against Windows, so M5's layouts only became trustworthy once a libavcodec
byte capture had verified them field by field. AV1 turns out not to need
that capture: `DXVA_PicParams_AV1` ships in the Windows SDK's OWN `dxva.h`
(10.0.26100.0 and 10.0.28000.0 on .173), which is the declaration the driver
was compiled against and therefore outranks any mirror.
So `layout-probe-av1.c` is committed beside pf-vaadec's probe, compiled with
MSVC against that header on .173, and every number below came out of it:
DXVA_PicParams_AV1 is 912 bytes with alignment 1, PicEntry 36, Tile 16, and
each offset is a compile-time assertion. The nested blocks are asserted
through their own types too, so a wrong internal layout cannot hide behind a
right outer one. Every assertion passed on the first build, which is the
result worth having: the transcription and the compiler agree.
The bit-field words are measured, not assumed. C bit-field allocation order
is ABI-defined rather than standardised, so the probe sets one member at a
time and prints the word, and the tests check each packer against what MSVC
produced — tx_mode at bits 22-23, reference_frame_update at 26, film grain's
sixteen-bit word with matrix_coeff_is_identity at 12, and so on.
Two places AV1 puts things where the other two codecs would not, both now
written down where a conversion will read them:
Global motion is per REFERENCE, inside DXVA_PicEntry_AV1 — where Vulkan
hangs one global-motion block off the picture info. A conversion carrying
the Vulkan shape across would leave every warped reference at identity.
CDEF strengths are packed two-to-a-byte, primary in the low six bits and
secondary in the top two, where both the AV1 syntax and Vulkan's Std block
keep parallel arrays.
A zeroed block names NO reference: 0 is a valid surface index, so a
memset-style default would quietly point every unused reference at surface
0, which decodes, and decodes wrong.
Gates: macOS fmt/clippy/347 tests, container clippy -D warnings over six
crates, 802 tests, workspace check.
The sequence header and the picture info, converted for
VK_KHR_video_decode_av1. Same shape as the H.264 and H.265 conversions, and
the same ownership contract: boxed backing beside the Std struct that points
at it, movable wrapper, no mutation, not Clone.
AV1 puts almost the whole frame header in the PICTURE info rather than in a
parameter set, so StdVideoDecodeAV1PictureInfo carries eight pointers to
per-frame blocks — tile info, quantisation, segmentation, loop filter, CDEF,
loop restoration, global motion, film grain — and the tile info carries four
more arrays of its own. Session parameters, by contrast, hold exactly one
sequence header. That asymmetry is why params_av1 is the small module here
and pic_av1 the large one.
The plan now carries the parsed frame header whole. The client needs a
digest — size, depth, colour, keyframe — but a backend needs nearly all of
the header, so AuPlan carries it the way its H.264 and H.265 siblings carry
their activated parameter sets: a backend builds from exactly what was
parsed, never by re-reading the access unit.
referenceNameSlotIndices holds DPB SLOT indices, not positions in the
reference list, and that is the HEVC RPS defect's exact shape in a narrower
place. Measured rather than argued: over the vendored vector the two
readings disagree 566 times across 274 frames, and the test fails if they
ever stop disagreeing, because then it would no longer be able to tell the
conventions apart.
Two places where transcription would have been wrong, both caught by the
types and then by asking the spec:
The parser's film-grain point arrays are 16 entries where the Std ones are
14 (luma) and 10 (chroma) — the spec's own maxima. The counts are validated
against the Std capacity and the copy is bounded by them; a stream declaring
more is refused, because a decoder handed fewer scaling points than the
stream declared synthesises different grain.
`coded_denom` is the superres denominator less SUPERRES_DENOM_MIN and only
meaningful where superres is in use, and `UsesLr` is derived — no frame
header codes it — from whether any plane's restoration type is not NONE.
Film grain rides only where the sequence enables it AND the frame applies
it, with the apply_grain flag set from whether a block is attached, so the
flag and the pointer cannot disagree.
Gates: macOS fmt/clippy/345 tests, container clippy -D warnings over six
crates, 800 tests, workspace check.
The third planner in this crate, and the foundation every AV1 rung will
consume. Same contract as its H.264 and H.265 siblings: an access unit in, a
plan out, with the vendored cros-codecs parser reading the bitstream and this
module owning the reference ledger, the output bookkeeping and the
concealment posture.
AV1's reference model is simpler than H.264's and entirely explicit — eight
numbered slots, `ref_frame_idx` naming what a frame reads and
`refresh_frame_flags` naming what it writes — so the planner is bookkeeping
rather than derivation, and a frame naming an empty slot is a lost reference
with no spec process that might legitimately have emptied it.
Two things measurement changed, both before a line of backend code depends on
them.
`plan_au` returns a VECTOR. An AV1 temporal unit may carry several frames,
and the vendored vector does: 250 units, 274 frames, 24 units carrying two.
Measured, those 24 extras are not `show_existing_frame` (there are none in
this vector) but HIDDEN frames — decoded, never displayed, referenced later.
A planner that took the last header in each unit would have decoded 250
frames and silently dropped 24 REFERENCES, and the damage would have
surfaced as missing-reference concealment on frames that were never damaged.
A picture is not removed until its LAST slot goes. One picture routinely
occupies several slots at once — a key frame refreshes all eight — so a slot
being overwritten does not mean its picture is gone. Reporting it removed
would free a surface under a live reference, which is precisely the shape
this program exists to catch. Tested directly, and asserted to report once
rather than once per slot.
What this does not cover is written down rather than left to be assumed: the
vector uses `show_existing_frame` zero times, so the display-only path and
its key-frame slot reset are exercised by no test here, and the test asserts
that count is zero so the day it changes the claim gets revisited.
Per-backend conversions are deliberately absent. Vulkan, DXVA and libva
disagree about what a reference list indexes — the disagreement that made
HEVC unplayable on every driver — so each belongs beside its siblings in
pf-vkdecode / pf-dxvadec / pf-vaadec, where its own convention is written
down and tested.
Gates: macOS fmt/clippy/344 tests, container clippy -D warnings over six
crates, 799 tests, workspace check.
The native VAAPI decoder now runs end to end: pf-vaadec's plans go into
libva's buffers, the surface comes back as DRM-PRIME dmabufs, and the
presenter imports them exactly as it does the FFmpeg rung's. Pin-only —
`PUNKTFUNK_DECODER=native-vaapi` — for the reason M5's D3D11VA rung was:
`auto` admission is earned with hardware parity and a soak, and this rung
has decoded nothing yet.
libva is dlopen'd rather than linked, so the pf-lxcheck2 container compiles
and clippies the whole thing without libva-dev, and a machine without a
VAAPI runtime gets a clean refusal instead of a packaging dependency.
The surface pool is not the slot map. `SlotMap::assign` hands out the lowest
free slot, and a slot freed by an access unit's own removals is free by the
time that unit's picture takes it — measured at 225 of the vendored vector's
250 access units. A surface bound by slot index would therefore decode, on
nine frames in ten, into the surface still holding the picture on screen. So
`plan_to_va` now takes the decode target as a parameter, bound by the caller
at activation time the way pf-vkdecode binds a pool image, and a surface is
free only when no live picture is bound to it, no output is owed for it, and
no consumer holds it.
Measured rather than transcribed, as everywhere else here: layout-probe.c
grew the export descriptor (312 bytes, objects[4]/layers[4]), the buffer-type
enumerators — VASliceParameterBufferType is 4 and VASliceDataBufferType is 5,
not the 3 and 4 that counting off the header suggests — and the config,
attribute and generic-value layouts. All pinned as compile-time assertions,
which is how the 12-byte VAGenericValue in the first draft was caught: the C
union holds a pointer, so it is 8-aligned and 16 bytes.
The plane walk lives in pf-vaadec, pure and unit-tested on macOS, because it
is the one structure the DRIVER writes and we read: SEPARATE_LAYERS returns
NV12 as two layers, and taking layers[0] is the green screen this project has
already paid for. It also refuses what it cannot express rather than guessing
— a bogus object count, a plane naming an object that is not there, objects
disagreeing on tiling.
Own DecodedImage variant, same payload type. The physical hand-off is
identical to the FFmpeg rung's, so the presenter keeps ONE arm and one
demotion streak; the variant exists so the compiler asks which rung decoded
wherever that matters. Both D3D11VA rungs share a variant and `1573a987` had
to fix the consequence afterwards — a "native" soak that could silently have
been an FFmpeg soak. Here the four uncovered matches were compile errors.
Buffers are destroyed by us, not by vaEndPicture: va.h is explicit that the
user must call vaDestroyBuffer, and the libva 0.x behaviour is long gone.
Leaking two per picture at 60 fps exhausts the driver's store in minutes.
pf-vaadec's presenter headroom was 4, written against no consumer. The Vulkan
rung had already measured the client pipeline at four to seven held frames;
it is 8 now, pinned to that crate's constant so a re-measurement moves both.
Gates: macOS fmt/clippy/341 tests/cargo doc, and in the container clippy
-D warnings over six crates, 795 tests, workspace check.
Hardware legs are still owed — no AMD/Mesa or Intel box was reachable.
The H.265 twin of plan_to_va, and with it pf-vaadec covers both codecs end to
end from an AuPlan to the buffers a vaRenderPicture call carries. What remains
for the rung is the Linux-only plumbing.
HEVC differs from H.264 in four ways that each had to be got right rather than
assumed, and they are why this is a separate module instead of a parameter:
ReferenceFrames is 15 entries, not 16.
The reference sets are FLAGS, not arrays. There is no RefPicSetStCurrBefore
here: membership is ORed into each DPB entry's own flags. Vulkan wants slot
indices in identically named arrays, DXVA wants list positions in them, and
VAAPI wants neither — three spellings of one idea, and confusing the first two
is what made HEVC unplayable on every driver.
The per-slice lists are INDICES into ReferenceFrames, not pictures and not
surfaces. So the DPB array is built first and every list entry resolved
through it; a picture a slice names that is not in the marked DPB is a refusal
rather than something to paper over, because there is nothing to fall back to.
The offset is in BYTES. slice_data() is byte-aligned by byte_alignment(), so
header_bit_size / 8 is exact — and a header that is not a whole number of
bytes is an error rather than a rounded offset, which would decode garbage
from the first inter picture.
Two conversions that are NOT copies, and would have been silently wrong as
copies: libva takes the derived ChromaOffsetLX (equation 7-56) where the
parser stores the coded delta, so putting the delta there would tint every
weighted-predicted block; and only 32x32 matrixIds 0 and 3 exist, where the
parser keeps six slots. The IQ matrix is Optional and gated on
scaling_list_enabled_flag for the reason review round 13 found on the DXVA
side — a driver MUST apply what it is handed, so a table of parser defaults
dequantises every residual to zero.
The weight table is only filled where 7.3.6.1 says one is coded, and the
chroma denominator is clamped into a legal shift so a malformed stream cannot
panic a decode thread.
Tests walk both HEVC vectors — the 250-frame 8-bit one and the 50-frame
Main 10 one, so a depth field wired to a constant would show — asserting per
slice that the start code was trimmed, the byte offset is inside the slice,
and every used list index points at a DPB entry that is actually valid. Per
picture it asserts that exactly the three current sets carry RPS flags and
nothing else does, and the walk fails if it never saw an RPS flag or a
reference at all, so it cannot pass vacuously.
The HEVC twin of pf-vaadec's H.264 buffer layouts, measured the same way: the
committed probe extended to cover va_dec_hevc.h, every size and offset read
off real libva 2.23.0 headers and pinned as const assertions —
VAPictureHEVC 28, VAPictureParameterBufferHEVC 604,
VASliceParameterBufferHEVC 264, VAIQMatrixBufferHEVC 1016 — and every
bit-field position read back out of a real header rather than counted by eye.
The finding worth carrying: HEVC's reference plumbing is a THIRD convention,
and this program has now been bitten by confusing two of them.
Vulkan takes DPB SLOT indices in RefPicSetStCurrBefore/After/LtCurr.
Writing reference-list positions there is what made HEVC unplayable on every
driver until it was root-caused.
DXVA takes positions into RefPicList[] in identically named arrays.
VAAPI takes neither. It marks set membership as FLAGS on the DPB entries
themselves — VA_PICTURE_HEVC_RPS_ST_CURR_BEFORE / _AFTER / _LT_CURR — and
its per-slice RefPicList[2][15] holds INDICES INTO ReferenceFrames, not
pictures and not surfaces.
Three spellings of one idea, identical names on two of them, and different
referents on all three. The conversion will say which it is writing, every
time, and the docs now hold all three side by side.
Two more asymmetries with the H.264 side, recorded where they will be read:
ReferenceFrames is 15 entries here, not 16; and the offset is
slice_data_byte_offset — BYTES, where H.264 wants bits — over the same
definition. slice_data() is byte-aligned by byte_alignment(), so the parser's
header_bit_size / 8 is exact rather than rounded, which the conversion will
assert rather than assume.
Tests cover the probe's measured bit patterns plus a disjointness sweep over
every field of pic_fields and slice_parsing_fields — two probe vectors per
word would not catch a shift typo that overlapped two neighbours, and these
words are 20 and 14 fields wide.
The second half of pf-vaadec: picture parameters, inverse-quantization
matrices and one slice-parameter record per slice, over the same transaction
discipline pf-dxvadec uses — validate, resolve references against the
PRE-removal slot map, then apply removals and assign the setup slot last. A
half-applied DPB update is the shape of a corrupt reference, so nothing
mutates until every fallible step has passed.
Three things VAAPI wants that neither other backend does, all of which the
existing plan already carries:
A bit offset. slice_data_bit_offset is where slice_data() begins, counted from
and including the NAL header byte with emulation-prevention bytes removed —
DXVA takes a byte offset, Vulkan takes nothing. It costs no new parsing: the
vendored parser records exactly that as SliceHeader::header_bit_size, because
cros-codecs' own production backend is VAAPI.
The slice data without its start code, since that offset is relative to the
NAL header byte. SlicePlan::data is start-code-inclusive and the prefix is
three OR four bytes — the host emits four on every access unit — so it is
measured per slice rather than assumed. Assuming it is the defect that made
HEVC unplayable on every driver.
The per-slice reference lists. DXVA's short-format slice control expresses no
lists at all; VAAPI wants RefPicList0/1 in 8.2.4.2 order, which is what the
plan's derived lists already are.
And the distinction that cost M5 a defect, now written down in a third place:
reference_frames is documented "in DPB", the same statement DXVA's
RefFrameList makes and the opposite of Vulkan's pReferenceSlots. It is filled
from the marked-DPB snapshot; the per-slice lists come from the slice's own.
Getting that backwards loses a long-term reference no slice happens to name.
Weight tables follow 7.3.3's presence rule rather than being copied
unconditionally: flagged only where the PPS actually enables explicit
weighting for that slice type and list. Flagging them otherwise hands the
driver defaults as though the stream had coded them. The vendored
PredWeightTable stores luma_offset_l0 as [i8; 32] but luma_offset_l1 as
[i16; 32] — an upstream inconsistency, not a semantic one — so the narrow side
widens.
Envelope refusals are errors, never silent narrowings: slice groups, separate
colour planes, a capacity mismatch, a reference holding no slot, lists past
their array bounds, a slice range outside its access unit.
Tests: 15. The one that matters walks all 250 access units of the vendored
conformance vector through H264Planner and this conversion, asserting per
slice that the range lies inside its access unit, that the declared size
matches it, that the start code really was trimmed, and that the header
neither is zero bits nor outruns the slice — plus that reference_frames
carries exactly as many valid entries as the marked DPB and every entry past
it is invalidated. It also asserts it saw a multi-slice picture and a
non-empty reference set, so a splitter bug cannot make it vacuous. Gates:
rustfmt, clippy, cargo doc with no unresolved links, and the container's
clippy -D warnings, tests and workspace check.
The companion to the Vulkan ten-bit leg, over the same vector and the same
P010 goldens — one golden file serves both rungs because a D3D11 P010 surface
and Vulkan's 3PACK16 family hold the ten bits in the same place.
This is the rung where the gap mattered most. D3D11VA exposes no per-picture
status query at all, so its HDR evidence was a session that built a Main10
decoder and streamed without complaint — which is precisely what a Main10
stream decoding to garbage would also produce. Now there is a number.
It exercises geometry the eight-bit legs cannot reach: P010 samples are two
bytes, so a row is width * 2 rather than width, and HEVC's 128-line granule
pads a 240-line picture to a 256-line surface — so the chroma plane starts a
long way from where the display height alone would put it. Getting either
wrong is the smeared-rows failure this project has already paid for once, and
it would have looked like a decoder fault.
The run body now takes the stream format and the expected access-unit count
rather than assuming the eight-bit envelope and 250 frames.
A CPU guard pins the vector at ten bits — 4:2:0, both depths minus8 == 2,
320x240, 50 access units. A regenerated eight-bit vector would otherwise turn
this into a second run of the eight-bit path under a ten-bit name, passing,
because its goldens would have been regenerated with it.
Hardware: HEVC Main 10 50/50 bit-identical on the RTX 4090 and on the AMD
Radeon iGPU, alongside the unchanged eight-bit legs at 250/250 on both. With
the Vulkan leg's two drivers that is four independent drivers across two
rungs for the ten-bit path, where yesterday there were none.
Every golden set in this program was eight-bit. So the strongest thing anyone
could say about ten-bit decode was that a Main10 session BUILDS and streams
clean — which is not the same claim, and is exactly the shape of claim this
program has been burned by. A Main10 stream decoding to garbage logs just as
cleanly: HEVC Main10 on D3D11VA has no per-picture status query at all, and on
the Vulkan side the devices that matter report queryResultStatusSupport=false.
The HDR legs were measuring that the pipe ran, not that the pixels were right.
So: a Main10 vector and its goldens, and a ten-bit leg that runs them.
The vector is 50 frames of 320x240 HEVC Main 10 4:2:0 from libx265 — 48 KB,
generated by a command recorded in the golden file's header along with
everything else needed to regenerate it. The goldens come from libavcodec's
software decoder and were cross-checked between two independent builds on two
architectures (ffmpeg 8.1.1 Homebrew/macOS-arm64 and 8.0.1 Ubuntu/x86_64),
which agreed on all 50.
The goldens are P010, NOT yuv420p10le, and that distinction is the whole
reason this could have quietly gone wrong: P010 puts the ten bits in the HIGH
bits of each little-endian 16-bit word with the low six zeroed, which is what
a D3D11 P010 surface and Vulkan's G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16
both contain. Hashing LSB-aligned samples against MSB-aligned ones would fail
on every frame on every driver, for a reason that has nothing to do with
decoding. One golden file therefore serves both native rungs.
The readback is now depth-aware. Its only eight-bit assumption was the second
region's buffer_offset, which is a BYTE offset where the extents around it are
TEXELS — that plus the buffer size are the whole change, because
bufferRowLength = 0 already packs rows at the copy extent. The expected pool
format moved onto the readback so the sizing and the per-frame assertion come
from one source; a readback sized for eight bits that then accepted a ten-bit
frame would hash half a picture and blame the decoder.
A CPU guard asserts the vector really is ten-bit — 4:2:0, both depths
minus8 == 2, 320x240, 50 access units, 50 planned outputs. Without it a
regenerated eight-bit vector would turn the ten-bit leg into a second run of
the eight-bit path wearing a ten-bit name, and it would PASS, because its
goldens would have been regenerated alongside it. That guard is not ignored,
so it runs on macOS and in the container rather than only on the fleet.
Hardware: HEVC Main 10 50/50 bit-identical on NVIDIA 610.43.03 (Linux) and on
the Steam Deck's RADV/VanGogh — first run on both, which also confirms the
P010/3PACK16 layout match rather than assuming it. The four eight-bit legs are
unchanged and still green on both boxes.
The VAAPI rung's crate, in the shape the other two native rungs established:
everything that can be a pure decision or a pure conversion lives in a
cross-platform crate the ordinary gates run, and only the parts that genuinely
need a device stay behind a platform cfg. This lands the first half of that —
the buffer layouts and the decoder-creation decisions — with the conversion to
follow.
Route: minimal FFI rather than cros-libva. The plan of record permits either
("cros-libva (or minimal FFI)"), and hand-declaring keeps the crate building
and testing on macOS and in the Linux container, which is the property that
made pf-dxvadec's defects findable on a laptop instead of on a box.
The layouts are not eyeballed. A C probe compiled against real libva 2.23.0
headers printed sizeof/alignof/offsetof for every field and set individual
bit-fields to read the resulting word back; those numbers are pinned as const
assertions, so a transcription slip is a compile error rather than a driver
reading the wrong byte. The probe is committed beside them, with the command
that runs it, because evidence that cannot be re-run is a claim.
What the probe settled that a reader would otherwise get wrong: VAPictureH264
is 36 bytes and is embedded 81 times across the two buffers, so its size is
load-bearing for every later offset; the three DEPRECATED FMO fields still
occupy bytes 624..628, and dropping them would shift everything after; and C
bit-fields allocate from the least significant bit on this ABI — proven, since
that is ABI-defined rather than standardised.
Groundwork for the conversion, established here so the next work package
starts from facts:
slice_data_bit_offset needs no new parsing. VAAPI is the only backend that
wants a bit position — DXVA takes a byte offset, Vulkan takes none — and the
vendored parser already records exactly it as SliceHeader::header_bit_size,
computed as (nalu.size - epb) * 8 - bits_left: from and including the NAL
header byte, emulation-prevention bytes removed. That is the field's
definition verbatim, and it is there because cros-codecs' own production
backend is VAAPI.
The slice data buffer starts at the NAL header byte, so the start code is
skipped — SlicePlan::data is start-code-inclusive and the prefix is three OR
four bytes, the host emitting four on 100% of access units.
reference_frames is the marked DPB, the same statement DXVA's RefFrameList
makes, so it comes from the dpb_refs snapshot; Vulkan's pReferenceSlots is the
opposite and takes the access unit's own set. All three conventions now have a
written home, which is the distinction that cost M5 a defect.
Unlike DXVA short-format, VAAPI wants the per-slice reference lists and the
full prediction weight tables inline — hence the 3128-byte slice record. One
wrinkle recorded rather than left to be discovered: the vendored
PredWeightTable stores luma_offset_l0 as [i8; 32] but luma_offset_l1 as
[i16; 32], and libva wants i16 for both.
Profile selection resolves H.264 to High for every 8-bit 4:2:0 stream instead
of reading profile_idc, because High is a superset for the tools our hosts
emit and picking Main for a stream that turns out to use 8x8 transforms is a
mid-stream failure where picking High is not. 4:4:4 and 10-bit H.264 are
refused rather than narrowed to an 8-bit profile — that class of silent
narrowing decodes to garbage instead of failing.
11 tests: the probe's bit patterns, a disjointness check per bit-field word
(two probe vectors alone would not catch a shift typo that overlapped two
fields), and the envelope refusals. Gates: rustfmt, clippy, cargo doc with no
unresolved links, and the Linux container's clippy -D warnings, tests and
workspace check.
The `stats:` line's decode-path tag is derived from the DecodedImage variant,
and both D3D11VA rungs deliver DecodedImage::D3d11 — they share the hand-off
ring on purpose. So a native-d3d11va session and an FFmpeg-d3d11va session
emitted a byte-identical tag, and nothing downstream could tell them apart.
The native Vulkan rung never had this: it carries its own variant, hence its
own `native-vulkan` tag.
That is not cosmetic, and it was found the only way it could be — by running
the rung on glass and having to grep the log to prove which one had answered.
A native pin that fails to initialise falls through to the FFmpeg rung by
design; the line it then emits is exactly the line the native rung would have
emitted. M5's owed soak and M9's vendor-matrix bake both rest on attributing
a session to a rung, and until now the machine-readable half of that evidence
could not do it. This project has already shipped one measurement that could
not tell "clean" from "unmeasured"; this is the same shape.
D3d11Frame now records which rung wrote the surface, keyed off the pin
constant itself rather than a second field the two rungs could set
inconsistently — the native rung passes DECODER_PIN into the hand-off and
nothing else does.
The stats line stays additive for every shipping session: the only value that
changes belongs to a rung that is pin-only and deliberately absent from the
automatic ladder, and the Windows shell passes the line through opaquely
rather than matching on the tag.
Verified on glass on .173, both directions: pinned native-d3d11va gives 85
windows tagged `native-d3d11va` and 0 plain, pinned d3d11va gives 64 plain and
0 native, zero errors either way. Gates: clippy -D warnings on Windows, the
Linux container's clippy/tests/workspace check, rustfmt.
The native D3D11VA rung had no pixel evidence at all. Its DXVA bytes were
checked against libavcodec's own captured bytes, and its Intel bring-up proved
the driver accepts the submission — but nothing had ever compared what came
out. This is that comparison, against the same goldens and the same reference
the Vulkan rung was held to: libavcodec's SOFTWARE decode, which is ground
truth rather than a peer implementation, so the two rungs' verdicts are now
directly comparable numbers.
It reads back the DECODE surface, before the VideoProcessorBlt, so what is
hashed is the half this rung is responsible for; the hand-off is the shared,
field-proven half and is deliberately not in the measurement.
Finding, recorded rather than papered over: this rung presents in DECODE
order. It never consults AuPlan::dpb.outputs — submit blits setup_slot and
returns. The native Vulkan rung keeps a display-order queue for exactly that
reason, and libavcodec's D3D11VA rung reorders internally, so this rung
differs from both. It cannot bite on punktfunk streams, which are zero-reorder
and carry no B pictures, but that is a convention of our hosts rather than a
structural guarantee, and a stream that did reorder would present out of order
with nothing to say so.
Both vendored vectors DO reorder — the H.265 one's first B picture at AU 3 is
what localised the RPS slot defect — so a harness hashing in decode order
would report a permutation against display-order goldens and read like a
decoder fault. Instead each decoded surface is hashed against the PicId the
planner gave it and the hashes are emitted in the planner's own output order.
The reordering is the test's, done by the planner the rung already trusts, and
`both_vendored_vectors_really_do_reorder` asserts the reason so the docs
cannot go stale silently.
The crop reads the chroma plane at RowPitch * texture height, not display
height: the decode pool is aligned to the codec's granule and is taller than
the picture. That is the 1088-row smear this project has already paid for.
Two CPU guards run in ordinary CI. This file needs its own Annex-B splitter
(pf-client-core does not depend on the vendored parser), and a splitter that
disagreed with pf-bitstream's would fail on hardware as a frame-count mismatch
that reads like a decoder defect; instead it fails on CPU, saying so.
PF_DXVA_ADAPTER pins a GPU by description substring and every run prints the
adapters it saw — .173 enumerates its AMD iGPU alongside the 4090, and which
one answered is a fact worth printing rather than inferring.
Hardware: H.264 and H.265 both 250/250 bit-identical on NVIDIA GeForce RTX
4090 and on the AMD Radeon iGPU, Windows. Gates: clippy -D warnings and the
lib tests on Windows, the Linux container's clippy/tests/workspace check, and
rustfmt.
Both vendored vectors carry three-byte Annex-B start codes throughout. The
real host emits four-byte ones on 100% of access units in both codecs —
1514/1514 H.264 and 1133/1133 HEVC, measured off the M0 NVENC corpus through
the capture hook's own .idx offsets. So every parity verdict this program has
recorded was taken on a prefix form that never ships, and the one form that
does ship was exercised by nothing.
That gap is not hypothetical. Submitting four-byte start codes to
vkCmdDecodeVideoKHR unchanged is exactly what made HEVC unplayable on every
driver tested: drivers are validated on the three-byte form, and a fixed
+3 + 2 skip into a four-byte-prefixed slice reads a nonsense pps_id — the
115 and 119 both NVIDIAs printed. H.264 was never safe here by structure,
only by its vendored encoder's convention, which is why the cure lives in the
shared ring layer and why this coverage is generic over both codecs.
Each codec's parity body now takes its access units as a parameter and runs
twice: once over the vector as it sits, once over the same vector rewritten
to four-byte prefixes. Prefix width carries no information, so both runs must
reproduce the same goldens — sharing one body is what makes that an equality
rather than two assertions that can drift.
The rewrite copies nalu.data[nalu.offset..], the same nal_size bytes the
parser hands the planner, so trailing_zero_8bits are dropped exactly where
the production parser drops them: the only difference between the two streams
is the width of every prefix.
Two CPU guards keep the new legs from passing vacuously, which is the failure
mode they are most exposed to — a rewrite that quietly returned its input
would make them trivially green and nothing on the fleet would notice. They
assert the original really does carry three-byte prefixes, that the rewritten
stream carries none, that the NAL count is preserved exactly, and that the
planner still yields 250 pictures.
Hardware: all four legs 250/250 bit-identical to libavcodec on two
independent driver stacks — AMD VanGogh on RADV/Mesa 26.0-devel (the Steam
Deck) and NVIDIA 610.43.03 on Linux. NVIDIA is the family that rejected the
four-byte form outright, so it is the meaningful witness for this regression.
wrong slice offsets
Two independent defects, both in this crate. pf-bitstream is untouched — its
HEVC plans were sound all along, which the D3D11VA rung proves by rendering
correctly from the same AuPlans.
The corrupter: StdVideoDecodeH265PictureInfo's RefPicSetStCurrBefore,
StCurrAfter and LtCurr carry DPB SLOT indices. We wrote positions in the
reference list. libavcodec's Vulkan HEVC hwaccel — the implementation every
driver is validated against — writes the index into its own DPB array and
passes that same value as slotIndex, while packing pReferenceSlots densely
over the used entries; the two numberings are provably different there, and
the RPS arrays follow the slot.
The two readings coincide on a freshly anchored stream, because the references
then occupy slots 0..n in reference-list order. They first diverge at the
vendored vector's first B picture, AU 3, where refs are slots 0, 2, 1 — so we
named slots 1 and 2 where the picture wanted 2 and 1, and every later access
unit inherited the error through its own references. That is why this shipped
and why review could not see it: correct for the opening pictures, wrong from
the first reordering onward.
It also accounts for the measurements exactly. Display order maps to decode
order as display 0 from AU 0, display 3 from AU 1, display 2 from AU 2,
display 1 from AU 3 — so the three frames that matched on AMD are precisely
the three access units where positions and slots agree, and 250 - 3 = 247 is
the divergence count that was measured. From AU 6 the named slots stop being
merely wrong and become unbindable by that operation, which is where NVIDIA
stopped reporting a verdict at all.
The diagnostic: every slice offset must point at a THREE-byte start code.
libavcodec discards the stream's prefix and writes 00 00 01, so that is the
only pattern drivers are validated on, and pf-dxvadec's packer already
normalised for exactly this reason and said so in its docs. This path uploaded
the prefix verbatim. 249 of 250 HEVC slice segments in the vendored vector
carry a four-byte prefix; all 500 H.264 slices carry three.
That derives the driver's own complaint bit for bit. A decoder reaching the
slice header by a fixed skip lands on the NAL header's second byte, reads
first_slice_segment_in_pic_flag as 0, and then takes six bits of the real
slice header as the tail of a long ue(v): 0xd0 gives 115, 0xe0 gives 119. A
P-slice header and a B-slice header — which is why exactly two bogus pps_id
values ever appeared.
⚠ H.264 was NOT protected structurally, only by its encoder's convention, and
the real host does not share that convention: every one of 1514 H.264 access
units and 1133 HEVC access units captured from an NVENC host prefixes its
slices with FOUR bytes. The vendored H.264 vector is therefore not
representative of what ships, and its bit-exactness was passing on a prefix
form the field never sends. The normalisation lives in the shared ring layer
and covers both codecs for that reason.
rebased_offsets is replaced by pack_slices, which trims the leading zero byte
and computes the offsets from the trimmed lengths in one call, so the bytes and
the offsets cannot drift apart; upload and the CPU test go through the same
pack_into.
Hardware, after the fix — H.264 AND H.265 both 250/250 bit-identical to
libavcodec, all four smoke legs green:
NVIDIA RTX 4090 610.88 Windows coincide
AMD Adrenalin 25.10.30.02 Windows distinct
NVIDIA RTX 5070 Ti 610.43.03 Linux coincide
On glass on the 4090 against a real NVENC host, 2800x1260 HEVC through the
auto ladder: 73 one-second windows all native-vulkan, fps avg 59.3 of 60,
decode 1.1 ms, e2e 4.5 ms p50, and ZERO driver-reported status failures where
the same session before the fix logged 1489 in 181 seconds and had dragged ABR
down to a 5 Mb/s target. No refusals, demotions, PlanWarnings, concealment,
DEVICE_LOSTs or panics.
Both defects now have CPU tests that were confirmed FAILING before the fix:
one walks every access unit of both vendored vectors and asserts each declared
offset opens on a three-byte start code, its own NAL header and a
first_slice_segment_in_pic_flag consistent with the segment index; the other
resolves every RPS entry by slot and asserts that 247 access units disagree
with the positional reading, so it cannot go vacuous on a stream where the two
happen to agree.
M3 was recorded as code complete. Its exit criteria named the HEVC gpu_smoke
and gpu_parity legs, and the goldens for them were committed — 250 per-frame
NV12 hashes, cross-checked between two independent FFmpeg builds, with a
header saying they are "consumed the same way by the HEVC parity test". No
such test existed. Both GPU files were H.264 only, with zero references to
h265, so nothing had ever decoded a single HEVC frame through this crate on
hardware.
They exist now, and the first run answered. On AMD Adrenalin 25.10.30.02
(distinct mode, queryResultStatusSupport=false) 247 of 250 frames diverge
from libavcodec, and that device's smoke leg PASSES — because smoke only
reads the driver's verdict and that driver reports none. That is the Ally X
class, reproduced in-house on demand: output that is wrong everywhere the
picture is looked at and clean everywhere the decoder is asked. Both NVIDIA
drivers reject the stream outright and name the cause themselves,
"Invalid PPS/SPS id in slice header (pps_id=119 / 115)" — the identical two
values, and the smoke leg dies at the identical AU 9, on a 4090 under 610.88
on Windows and on an RTX 5070 Ti under 610.43.03 on Linux. Same wrong values,
same access unit, two GPU generations, two operating systems: deterministic,
and therefore ours rather than any driver's.
It is not an ordering fault. Five of the divergent hashes appear nowhere in
the 250 goldens, so the pixels are wrong rather than correct-but-reordered.
Parity dies at frame 1 while smoke dies at AU 9 only because smoke holds four
frames before it looks; the first inter-predicted picture is already corrupt.
The legs are committed ahead of the fix deliberately. They are the regression
test for the defect, they are #[ignore]d so no CI leg changes colour, and the
evidence above is worth recording in the order it was obtained.
Adding a third and fourth copy of ~150 lines of unsafe Vulkan bring-up was
not acceptable, so it moved to tests/common. The two behavioural differences
between the callers are now named parameters rather than accidents: the parity
legs read back on a graphics queue and require one, while the smoke legs
accept a decode-only device and fall back to the decode family — which also
decides whether pool images are EXCLUSIVE or CONCURRENT, so it is load-bearing
rather than cosmetic. H.264 came through the refactor unchanged, verified two
ways: argument-by-argument against the previous file, and on hardware, still
250/250 bit-identical on NVIDIA Windows, AMD Windows and now NVIDIA Linux.
The loader is deliberately leaked at teardown. ash::Entry owns the Arc<Library>,
so dropping it unloads the Vulkan loader with every ICD and implicit layer;
harmless while each binary held one GPU leg, but each now holds two, and the
second would re-open a loader the first had torn down.
Three guards run without a GPU, because everything above is #[ignore]d: the
golden file's count and digest shape, the HEVC access-unit split agreeing with
what the CPU planner emits (with iraps == 1 pinning "no CRA anywhere", so a
re-synced vector that opens with one fails here rather than as a frame-count
mismatch on the fleet), the vector staying Main 4:2:0 8-bit since both legs
hard-code that probe, and a refusal to run the smoke legs with
PF_VKD_TEST_READBACK set, which would quietly grow the pool a usage flag
production never carries.
The native D3D11VA rung was committed with zero hardware evidence and, more
to the point, zero evidence of any kind: every claim it made about the DXVA
structures rested on reading the specification and reading libavcodec, and
reading is exactly the method that produced the four defects the last review
found. Three of those lived where a smoke test cannot look — in the buffer
descriptors — so a passing session would have proved nothing about them.
So this compares our submission against libavcodec's, byte for byte, on the
same access units of the same two vendored vectors. The reference side comes
from an FFmpeg n8.1 cross-built for Windows with the DXVA paths instrumented
(the recipe is in the harness docs); the comparison covers picture parameters,
quantisation matrices AND the buffer descriptors, 250 AUs per codec:
H.264 / HEVC picture parameters 250 AUs, no undocumented divergence
H.264 / HEVC quantisation matrices 250 AUs, no undocumented divergence
H.264 / HEVC buffer descriptors 250 AUs, no undocumented divergence
It found a real defect immediately. The DXVA short slice record is TEN bytes
— dxva.h packs these bitstream-layout structs to a byte — and this crate
declared it `repr(C)`, which pads {u32,u32,u16} to twelve. libavcodec's own
descriptors say so twice over: 20 bytes of slice control for a two-slice
H.264 picture, 10 for a one-slice HEVC one. Record 0 survives either way
(its fields sit at 0/4/8 regardless), so the mistake is invisible on a
single-slice stream and displaces every later record by two bytes on a
multi-slice one — which punktfunk hosts emit. Both records are now
`repr(C, packed)`, and the HEVC slice-control test grew a second record
because one record is the shape that hid this.
The audit that followed matters more than the fix. Per-field offset asserts
cannot see TAIL padding, which is what this was, so all six hand-declared
structs now also assert that their size equals the last field's offset plus
that field's own size. Under that rule the slice records were the only place
packed and natural alignment disagree — 1040, 232, 224 and 1000 were right
all along, and now provably rather than luckily. The module docs claimed
`repr(C)` "reproduces MSVC's default packing exactly for that shape"; that
was a guess wearing a proof's clothes, and it is gone.
Two differences are documented rather than fixed, each with the argument for
why it is inert. libavcodec seeds prev_poc_msb = 1 << 16 at every IDR, so its
POCs are the specification's plus 65536 uniformly; every use a driver makes
of those fields is a difference, and references match on FrameNumList, so the
harness compares POCs relative to that constant and requires it to hold on
every AU rather than importing a magic number into a derivation the Vulkan
rung shares. And HEVC's loop_filter_across_tiles_enabled_flag is inferred 1
by 7.4.3.3.1 when the PPS codes no tiles while libav leaves it 0, with tiles
disabled either way. Both ride a channel that always prints, and both are
guarded by tests that synthesise the differences an allowance must NOT
absorb — a documented divergence that swallows a real defect would be worse
than no harness at all.
Everything checkable without a capture is now a non-ignored test: the buffer
set and order per codec, NumMBsInBuffer's codec asymmetry (mb_width*mb_height
on H.264's bitstream and slice-control buffers, zero everywhere for HEVC),
the three 7.4.5 scaling-list cases, contiguous slice records tiling DataSize,
the 128-byte padding charged to the last record and no other. That is the
part which would have caught the last round's defects with no hardware at all.
The Windows fallback rung, and auto's first choice on Intel, now has a
native implementation driven by pf-bitstream's plans instead of
libavcodec. New crate pf-dxvadec holds everything that can be a pure
function — the DXVA structure layouts, both codec conversions, bitstream
packing, config selection — deliberately CROSS-PLATFORM, because a
cfg(windows) module is verified by a remote cargo check and nothing
else, and this milestone's riskiest code is exactly the part no local
test can see. Only the FFI lives in video_d3d11_native.rs.
windows-rs does not generate dxva.h at the pinned rev, so the DXVA
structures are hand-declared: compile-time assertions on every struct
size AND every field offset, packed bitfield words as plain integers
with named builders and the bit positions written beside the C
declaration, and a const zeroed() per struct so construction needs no
unsafe at all. The crate's only unsafe is a sealed byte view over those
PODs. Review round 13 checked all seven layouts field by field in
declaration order — sizes, widths, array lengths, the PicEntry
index/flag packing, and every named bit's position and width.
The decode pool reproduces libavcodec's rather than inventing one:
ONE texture with ArraySize = pool size, BIND_DECODER and nothing else,
MiscFlags 0, aligned 16 for H.264 and 128 for HEVC. That is deliberate.
This rung's predecessor records that a hand-built pool which validated
on NVIDIA was rejected by Intel at the first SubmitDecoderBuffers — and
Intel is the vendor this rung exists for. The VideoProcessorBlt into
shareable RGBA is untouched: importing a multiplanar NV12 D3D11 texture
into Vulkan device-losts on NVIDIA, so that hand-off is load-bearing
field-proven code. It was extracted into a shared HandoffRing so both
rungs fill one implementation; the review diffed the blit statement by
statement, including the keyed-mutex pairing.
Review round 13's four defects are fixed. The blocking one: the HEVC
quantisation matrix was submitted unconditionally, and the vendored
parser leaves it ALL ZEROS unless the stream codes one — unlike FFmpeg,
which seeds the spec defaults. On a stream saying 'use the default
matrices' the driver is obliged to apply what it is handed, so every
residual would dequantise to zero and the picture would drift to flat
prediction. It is now gated on scaling_list_enabled_flag exactly as
libav gates it, with the Table 7-5/7-6 defaults supplied when enabled
but uncoded.
Second: NumMBsInBuffer was 0 where libav's H.264 path sets
mb_width * mb_height. This module's whole method is verbatim
reproduction on precisely the call that once failed for Intel, so an
omitted descriptor field is the same class of bug as the pool.
Third, and the one to watch on hardware: RefFrameList carried the
frame's reference set rather than the pictures marked used for
reference. Vulkan defines pReferenceSlots as the slots this operation
uses, so a subset is correct there; DXVA defines RefFrameList as a
statement about the DPB. The list DERIVATION survives a subset — which
is exactly why a smoke test would have passed — but a long-term
reference held across frames that none of them name would vanish and
reappear, and a driver keeping per-reference state is entitled to
discard it in between. That is the Ally X symptom shape. pf-bitstream
now exposes a per-AU DPB snapshot for both codecs and the converters
build the array from it, frame references first, marked tail appended.
121 of the 250 vendored AUs carry a marked picture the frame never
names, so this is exercised, not theoretical.
Fourth: the session identity omitted bit depth and chroma, while the
Windows host flips an HDR desktop to PQ in-band with a new SPS — a
depth change at unchanged size would have decoded 10-bit samples into
an NV12 pool. Identity now derives from the SPS per AU and rebuilds.
Wired PIN-ONLY (PUNKTFUNK_DECODER=native-d3d11va), absent from every
auto arm. Nothing has decoded a frame yet, and M2's discipline was that
auto admission comes only after hardware parity. A runtime streak
demotes to the FFmpeg D3D11VA rung first, then software.
Also scaffolded: a byte-diff harness against libavcodec's own DXVA
picture parameters, with the FFmpeg patch and capture recipe in its
docs. Nothing here is checked against libav's actual bytes the way M3
was checked against its pixels, and that is the cheap way to buy the
confidence before hardware.
Gates: fmt clean; container clippy -D warnings zero across
pf-client-core + pf-presenter + pf-vkdecode + pf-dxvadec +
punktfunk-core; tests 73/131/63/129/354 green; cargo check --workspace
clean; Windows cargo check and clippy -D warnings clean on .173.
This program exists because a field corruption was architecturally
undetectable through FFmpeg: no decode-status read, no corrupt-frame
flag, errors only as scraped log lines, and no recovery-point signal so
intra-refresh healing was invisible. The native decoder has all of those.
M4 is where they stop being internal.
DecodeHealth counts, per session and without allocating per frame, what
the three answers actually are: damaged (the stream arrived incomplete),
refused (the rung would not decode it at all) and driver-failed (the
hardware says it could not decode what arrived), plus the current and
worst concealment run — the figures that separate one bad AU from a
stream that never came back. They ride the stats line additively, so an
FFmpeg session and a healthy native session emit byte-identical output
to today. The status-query capability is reported too: without it a
clean report cannot be told from an unmeasured one, which is the whole
nb_queries=0 lesson.
The headline is local recovery. Until now the pump could only learn that
intra-refresh healing finished from wire flags the host sends; absent
those it froze until the 500 ms backstop forced an IDR. The parsed
recovery-point SEI now feeds the re-anchor gate directly, so a session
lifts on the picture that is actually clean. Wire semantics are
untouched for every client that never calls it.
Detection now asks for recovery instead of erroring — an integrity
warning ticking the error streak would demote the native rung on exactly
the lossy links it exists to diagnose, where an FFmpeg rung conceals
silently and keeps its job.
Review round 12 found that trade had removed the escape hatch entirely.
Concealment returning Ok(None) reset the demotion streak, and worse: the
driver-verdict ledger is only populated when a frame ships, so under
continuous concealment no verdict was ever read and the erroring arm
could not fire at all. A host framing regression of the 0.23.0
slice-wire class — which does not self-heal, and which a keyframe does
not clear — would have frozen indefinitely with no demotion and a clean
integrity line, where before it demoted to FFmpeg-Vulkan and showed a
picture. Now only an answer that proves the rung works clears the
streak: a shipped frame, or a clean no-frame. Concealment neither ticks
nor clears, so a lossy link still cannot demote a healthy rung while a
driver failure interleaved with concealment reaches the threshold again.
Two more honesty defects from the same round. A rung refusing every AU
reported no integrity line at all — the founding failure mode, wearing
the shape of a clean bill of health; refusals are now counted. And
driver-failed could be non-zero on a device that cannot produce driver
verdicts, because a degraded timeline read looked the same as one; the
attribution is now withheld inside the counter rather than at call
sites, so the self-contradictory line is unrepresentable.
Local recovery also no longer trusts any recovery-point SEI: only one
whose target advances past an outstanding wave counts as a new wave, so
an encoder re-announcing the current wave with a decreasing count — legal,
and what x264 intra-refresh does — cannot lift the freeze early onto a
partially stale picture. Frames buffered across an arm are dropped by
decode order for the same reason.
Fault injection is a first-class tool now (PUNKTFUNK_AU_FAULT, inert
unless set, env read once). Its test replays the vendored vectors
through the real planners and asserts a negative the plan assumed away:
truncation and bit flips are PROVABLY invisible to the parser — Annex-B
carries no NALU length, so a cut slice is just a shorter slice and a
flipped payload byte is syntactically perfect. Only dropped AUs are
parser-detectable; the rest need the driver verdict, which is why the
status query matters. The H.265 leg found a second: three of that
vector's faulted AUs are sub-layer non-reference pictures, so dropping
them damages nothing and silence is correct — the test asserts both
verdicts and guards that neither half goes vacuous.
Per-frame decode latency was deliberately NOT built. Polling answers
only 'complete by now', and the pump polls once per AU, so every sample
would quantise up by as much as a frame interval — 8.3 ms at 120 Hz
against decodes of 0.1-2 ms. Sampling faster needs a spin or a second
thread on a decoder that is deliberately not Sync. A blocking per-frame
wait is the field scar that once capped a stream at 51 fps. The honest
sampled stat stands.
Also fixed, pre-existing: the re-anchor gate re-armed on every damaged
AU, so sustained damage permanently zeroed the mark count — meaning the
wire's two-mark rule could never complete on exactly the lossy links it
was written for.
Field note recorded while wiring this: intra_refresh_recovery is set by
exactly one encoder backend (Linux libav-NVENC under
PUNKTFUNK_INTRA_REFRESH). AMF and QSV run a wave with no wire mark, and
AMF emits no recovery-point SEI either, so AMD/Windows intra-refresh
sessions still have no clean recovery point by either route.
Gates: fmt clean; container clippy -D warnings zero across
pf-client-core + pf-presenter + pf-vkdecode + punktfunk-core; tests
69/131/129/354/41 plus 5 fault-detection green; cargo check --workspace
clean.
The last piece of M3 WP-2 — VkH265Decoder was built and hardware-gated
but nothing drove it. video_vk_native.rs holds a two-arm codec enum and
forwards to it; the ledger, release tokens, status-query settling and
timeline waits are byte-for-byte what they were, since they were always
codec-agnostic over one DecodedVkFrame contract. The forwarders are
written out per arm rather than macro'd so the unchanged H.264 arm is
visible to a reviewer.
The picture's own format now reaches the presenter, which picks bit
depth and MSB packing from it instead of assuming the H.264 envelope.
That incidentally fixes a live bug on the SHIPPING FFmpeg-Vulkan path:
it derived ten-bit-ness by comparing against the 10-bit 4:2:0 format
alone, so a 10-bit two-plane 4:4:4 surface — which its own format table
accepts, and which NVIDIA reports for HEVC RExt — got 8-bit range and
transfer maths. Reachable today with Full chroma plus 10-bit: decoded
correctly, displayed wrong.
Review round 11 caught a regression this WP would otherwise have
shipped. pf-vkdecode refuses a stream whose (chroma, depth) pair has no
picture format on the device, but the session is built lazily from the
first SPS, so the refusal arrived AFTER construction — past the point
where a native init failure falls through to FFmpeg-Vulkan. It burned
the error streak instead and demoted to VAAPI/D3D11VA, which on
NVIDIA/Linux means software. Turning on Full chroma on any non-NVIDIA
GPU was enough: a 4K HEVC session that ran on FFmpeg-Vulkan before this
branch would have landed on software decode.
Both halves are fixed. The negotiated chroma and bit depth — already at
the call site, the PyroWave arm four lines up uses them — are threaded
into the backend, which probes the same caps path ensure_state would
run, so the whole class refuses at CONSTRUCTION where the fall-through
already exists. For the legs no negotiation can carry (a level above
maxLevelIdc, an SPS that disagrees with the Welcome) the decoder latches
'never delivered a frame' and routes that first streak to FFmpeg-Vulkan
rather than down the hardware ladder. H.264 is deliberately not probed:
its envelope is fixed, so a probe would only add a profile guess on the
bit-exact path; it gets the latch as its backstop.
Two more from the round. Planner warnings are typed again rather than
Debug strings — pf-vkdecode simply lacked the h265 re-export its h264
twin already had — which restores the H.264 log rendering exactly and
unblocks M4, whose job is counting concealment by kind. And concealment
is now the integrity set only: NonZeroReorder is documented spec-legal
and fully planned, but the client treated every warning as damage, so
the opening IDR and every ABR renegotiation's IDR were released unshown
and re-anchored — a visible hitch on a healthy stream.
Also: a raw-format newtype so a neighbouring i32 field cannot be passed
to the colour maths, the presenter's depth table now pinned against
pf-vkdecode's actual output vocabulary rather than the FFmpeg lane's,
a per-format warn latch, and four stale docs.
Gates: fmt clean; container clippy -D warnings zero across
pf-client-core + pf-presenter + pf-vkdecode; tests 69/125/108/40 green;
cargo check --workspace clean.
M3 WP-2 complete. caps_h265.rs builds the profile the stream actually
needs (profile idc + chroma + bit depths, all three stated on every
Vulkan object) and resolves its picture format — Main to NV12, Main 10
to P010, RExt 4:4:4 to the two-plane 4:4:4 formats — validating it
against the format list of every role the chosen arrangement creates
images in. A Main 10 stream on an 8-bit-only device is refused BEFORE a
session exists, never narrowed: decoding 10-bit into an 8-bit surface is
the silent-wrongness class this crate exists to refuse. session_h265.rs
adds the three-array parameters ledger; decoder_h265.rs adds
VkH265Decoder, mirroring VkH264Decoder method-for-method so the client
wiring is a two-arm dispatch away.
H.264 and H.265 now SHARE the machinery instead of duplicating it:
derive_arrangement (one coincide/distinct/layered decision table),
ring::rebased_offsets (the slices-only rebase — non-VCL NALUs in the
decode range hang VCN firmware), session::bind_session_memory, and a
parameterised build_frame. A DecodeProfile enum replaces the bare
profile idc that images.rs and ring.rs used to take: both codecs' idc
types are c_uint, so handing an H.265 idc to the H.264 path COMPILED
SILENTLY and built a mismatched profile chain. That is now
unrepresentable.
The VPS leg is the ledger's real work. The vendored parser attaches a
VPS to an SPS only when it saw the NALU, and clients join live streams,
so VpsSource is Parsed-or-FromSps and is stored BY VALUE: re-activating
a VPS-less SPS is Current (no churn), but the real VPS arriving under
the same id is a content change and RECREATES onto it, because Vulkan
cannot replace a stored parameter set.
Review round 10 (adversarial) confirmed the hardware-proven H.264 path
is NOT regressed — derive_arrangement's check order and error identity
are byte-for-byte the original, build_frame's call sites still pass the
granularity-aligned extent (the 1088-row scar stays shut), and
rebased_offsets reproduces the deleted inline loop for every input while
moving the sum to u64 so overflow errors instead of wrapping. Also
verified: the refs-order contract on every path, the RESULT_STATUS caps
gate (each of reset/begin/end individually gated, no pool created when
unsupported — recording one on RADV hangs its VCN), pNext lifetimes, and
that no panic is reachable on stream input.
Its 10 findings are fixed. The two that mattered:
- A failed decode stranded a DPB slot. Once plan_to_vk_h265 had mutated
the slot map, five later failure paths returned without restoring it,
so planner and slot map both believed a picture was resident while no
image held it — and every later AU referencing it failed, where H.264
soft-degrades and keeps delivering. Fail-closed is kept (substituting
a reference silently is the corruption-hiding this program exists to
end) but made RECOVERABLE: a latch flushes the planner to AwaitingIdr
and resets the bindings on the next decode, which composes with the
client already requesting a keyframe on every decode error. The fix
deliberately covers pre-mutation failures too — those strand the
picture the other way round and wedge identically.
- DecodedVkFrame carried no picture format, so a Main 10 frame would
decode correctly and be rendered with 8-bit transfer/range math. It
now carries one, stamped from the pool so it is truthful for both
decoders by construction. The presenter comment says depth 8 is
because only H.264 is WIRED, not a decoder limit.
Plus: bind_session_memory freed allocations before the session that may
hold them was destroyed (an ordering regression from the extraction,
with a SAFETY comment asserting the opposite) — the bind-stage exit now
hands them back so Drop destroys first; max_level_idc is codec-tagged
rather than an H.264 type carrying H.265 code points; and the decode
family's videoCodecOperations is now checked, turning 'create an H.265
session on a device without the extension' from UB into a clean ladder
demote.
Deferred by design: no HEVC gpu_smoke/gpu_parity yet (its goldens are
already in tests/data/test-25fps-h265.nv12.sha256), and no codec
dispatch in the client — both later legs.
Gates: fmt clean; mac clippy zero warnings, pf-vkdecode 106 +
pf-bitstream 69 green; container clippy -D warnings zero for
pf-client-core + pf-presenter + pf-vkdecode, tests 69/121/106 green.
HARDWARE (.173, after the refactor — review saying the proven path is
safe is not the GPU saying it): gpu_parity '250 frames bit-identical to
libavcodec software decode' on BOTH the NVIDIA 4090 (610.88, coincide
mode) and the AMD iGPU (Adrenalin 25.10.30.02, distinct mode), gpu_smoke
green on both. Two independent drivers, both DPB modes, still bit-exact.
The smoke trace also shows the new videoCodecOperations capture reading
DECODE_H264 | DECODE_H265 | DECODE_AV1 off the real decode family.
M0's capture hook has been in since 119ec0dd with nothing consuming its
output. corpus_replay.rs is that consumer: point PF_CORPUS at an
au-<stamp>.<codec> capture and every AU walks back through the H.264 or
H.265 planner, asserting no errors and no warnings — a clean capture of
a healthy session must plan whole. Ignored by default (captures are
hundreds of MB and live outside the repo).
It earns its keep immediately. Captured on .173 against the live host
(NVENC, 2800x1260, ~30 s each, client-side codec pin only — no host
config touched):
h265 1133/1133 AUs planned, 0 errors, 0 warnings
h264 1514/1514 AUs planned, 0 errors, 0 warnings
The HEVC number is the point: it is the FIRST validation of the WP-1
h265 planner against real host output rather than the vendored
conformance vectors, and it lands before the client's HEVC rung exists
to produce on-glass evidence.
Two real-capture facts the harness had to learn, both from this run:
ending a capture means killing the client, so the final .idx line is
routinely half-written and the final AU's bytes may not all have landed.
Both are tolerated at the TAIL only — a malformed line anywhere else, or
a gap the data cannot cover mid-file, still fails loudly rather than
silently replaying a subset.
Also adds tests/data/test-25fps-h265.nv12.sha256: 250 per-frame NV12
hashes of the vendored HEVC vector from libavcodec's software decoder,
cross-checked frame-for-frame between two independent FFmpeg builds
(8.0.1 in pf-lxcheck2, 8.1.1 from Homebrew) — the sibling of the H.264
goldens, ready for WP-2's parity leg.
Gates: fmt clean; pf-bitstream clippy clean, 69 tests green (the replay
stays ignored in normal runs).
main moved from 8983ec04 to 35ba64ca while this branch sat open, taking the
release from 98 commits to 135. Merged in and folded the new work into the notes.
The largest addition is a new `## Before you update` section, because this batch
carries changes that need the reader to DO something and they were not going to
survive being buried in a Fixed bullet:
* Linux users of the virtual Steam Deck pad must `usermod -aG punktfunk` and
log back in, or it stops attaching — the capability moved off the `input`
group (which every gamepad guide tells you to join) onto its own, because it
can emulate arbitrary USB hardware.
* Plugin UIs moved to their own origin on PORT+1, so a self-signed console
needs the new port trusted once, and custom firewalls/proxies need it opened.
* Saving a custom launch command re-confirms the console password, and add-ons
may no longer set launch/pre-launch commands at all — a real break for any
third-party add-on that populated them.
* A fresh install now runs the plugin runner by default (upgrades untouched).
* The Deck setup script used to leave the generated console password
world-readable, so rotating it is worth a sentence.
The library-sources work is written as GROUNDWORK, deliberately. All six built-in
scanners still ship, still on by default, and nothing is removed — and none of
the replacement add-ons are published yet, so the migration banner only appears
as they arrive. Promising a user they can move Steam to an add-on today would be
the v0.22.3 mistake again: notes describing a build nobody is getting.
Two other honesty items. The Android HUD entry says outright that the stream did
not get faster and the headline number only got smaller because it stopped
counting the compositor's wait — otherwise every reader takes it for a speed-up.
The Windows non-C: settings entry says plainly that nothing is recoverable,
because the writes never reached disk, so there is no orphaned copy to restore
and the reader has to re-enter their preferences once.
`56adb470` (pad-audio WASAPI module path) is deliberately NOT a user-facing Fixed
entry: verified it is not an ancestor of v0.24.0, so it repairs a Windows build
break in code that has never shipped. It folds into the pad-audio feature. Same
for `19f637ea`, which is CI-only.
Under the hood gained the origin-isolation mechanism, the allowlist authorization
gate that fails the build on an unclassified route, store claims and the v2
library.json shape, the registry auth work, the config-writer fallback, send
pacing, and the vendored Deck WSI layer. The unverified list grew too: the origin
split has not been in a real browser, the packaging default-on changes have had
no installer run, and no launcher tile has ever been clicked.
Re-verified after the merge, all green: lock diff versions-only 32/32 against
origin/main, `cargo metadata --locked` resolves (35 members), `cargo fmt
--all --check` clean in both workspaces, doc lazy-continuation scanner 0 hits
over 521 files, notes body 0 internal-vocabulary hits above `## Under the hood`,
Play notes still 494/500 by android.yml's own gate logic. Wire 2, C ABI 16, and
the capability bytes are all unchanged from the bump commit — host_caps still has
exactly one free bit (0x80).
Play's "What's new" is left as it stands: at 494/500 there is no room, and the
only Android-facing additions here (the stats-overlay measurement change and a
certificate-strictness fix) are both worth less to a phone user than any line
already in it.
PROVENANCE deviations #6 and #7 carried 'report upstream' — done:
- chromeos/cros-codecs#99: h264 PictureData display_resolution
double-counts the left/top crop and underflow-panics on
parser-valid crop offsets (answers their open #81).
- chromeos/cros-codecs#100: h265 parse_slice_header index-OOB panic
when num_long_term_sps + num_long_term_pics > 16 — a hostile-input
panic on the LTR path (an instance of their #78).
Both reports offer the downstream patch for the AOSP tree.
M3 WP-2, first half. params_h265.rs: VPS/SPS/PPS -> StdVideoH265* with
owned pointer-backing (the params.rs contract), scaling lists incl. the
32x32 two-matrix quirk and +8 DC convention, short-term RPS re-encoded
from the parser's RESOLVED DeltaPoc arrays back into delta_poc_sX_minus1
syntax under monotonicity checks, fallback_vps_from_sps for streams
whose VPS NALU was lost. pic_h265.rs: plan_to_vk_h265 — h265 AuPlan ->
StdVideoDecodeH265PictureInfo + per-reference infos; the binding set is
the union of the three current RPS sets with the Std index arrays
indexing into refs (0xFF unused; the GPU half must lay pReferenceSlots
out in refs order); NumDeltaPocsOfRefRpsIdx from the predicted-from
candidate; transactional SlotMap lifecycle identical to pic.rs. SlotMap
reused unmodified — HEVC's ceiling equals H.264's 16+1.
Envelope fails closed: Main/Main10/MainStill/RExt only, 4:2:0-8/10 +
4:4:4 only (separate_colour_plane_flag rejected — ChromaArrayType 0 in
disguise), SCC palette predictors out, >64 ST RPS sets / >16 per side /
>32 LT SPS candidates out, checked narrowing on every narrower Std
field. No panics on untrusted input.
Review round 9 (adversarial): RPS re-encode math, Std field-by-field
conformance, transactionality and slot ceiling verified clean; 6
findings fixed pre-commit. Headline (BLOCKING): long_term_ref_pics_
present_flag=1 with num=0 left pLongTermRefPicsSps NULL — the header
demands a valid pointer whenever the flag is set, and flag=1/num=0 is
exactly the punktfunk LTR/RFI recovery stream shape; the all-zero
backing now rides whenever the flag is set. Also: the slice_offsets doc
in BOTH pic modules claimed submit-as-planned while decoder.rs packs
slices-only and rebases (non-VCL NALUs in the decode range hang VCN
firmware) — reworded so the HEVC GPU half cannot implement the hang; a
concealment-produced ST/LT duplicate now ORs the long-term flag across
occurrences; NumDeltaPocs clamps became a typed error; dead
UnmappableLevelIdc variant dropped.
Deferred to the GPU half: HEVC caps/profile chain, session parameters
(VPS leg in the ledger), P010/4:4:4 pool selection, recording, and the
pReferenceSlots-in-refs-order contract consumption.
Gates: fmt clean; mac pf-vkdecode 80 + pf-bitstream 69 green, clippy
clean; container clippy -D warnings zero (pf-client-core, pf-presenter,
pf-vkdecode) + tests green (69/121/80).
Closes M2. The 2026-08-05 ladder decision: WP-D shut with bit-exact
parity vs libavcodec (250/250 AUs on RADV, AMD-proprietary and NVIDIA)
and a 92-minute clean soak, and the program's goal is dropping FFmpeg
from the client — so on H.264 sessions where caps pass, auto now tries
pf-vkdecode FIRST, exactly where the ladder would reach FFmpeg-Vulkan.
No bake period.
native_vulkan_gate widens from by-name-only to the auto family
(auto/""/hardware); the vendor-first rungs are untouched (Linux
Intel/unknown still VAAPI-first, Windows Intel/unknown still
D3D11VA-first — NVIDIA and ALL AMD go native first). A native INIT
failure or caps refusal logs and falls through to FFmpeg-Vulkan, so
admission can't cost a session its decoder at start; runtime error
streaks ride the existing demotion machinery unchanged (past
FFmpeg-Vulkan to VAAPI/D3D11VA/software — a native→FFmpeg-Vulkan
runtime rung is deliberately absent, FFmpeg is on its way out).
PUNKTFUNK_DECODER=native-vulkan stays as the explicit pin; vulkan
keeps naming the FFmpeg backend specifically. A native_tried guard
keeps a failed pin init from re-attempting construction in auto.
Review round 8 (adversarial): no blocking code defect — no demote
bounce-back (Decoder::new is session-start-only; demotion mutates in
place), no double attempt, no cfg imbalance. 5 findings fixed: two doc
overclaims ("nothing regresses" now scoped to init; the ladder
enumerations no longer claim desktop-AMD Linux is VAAPI-first —
prefer_vulkan_first is vendor-wide), stale opt-in claims in Cargo.toml,
stale user-facing ladder text (console-ui row, trust.rs decoder field,
session README incl. the env-knob list), and the gate test now pins the
H264 codec-op bit to the literal 0x1 so a typo'd constant can't make
native silently never engage.
Gates: fmt clean; container clippy -D warnings zero for pf-client-core +
pf-presenter + pf-vkdecode; container tests green (pf-client-core lib +
pf-vkdecode + pf-bitstream); pf-console-ui check clean; mac
pf-vkdecode/pf-bitstream/cros-codecs 167 tests green.
On-glass sanity CLOSED 2026-08-05 ~22:10 UTC on .173 (4090, coincide
mode), decoder=auto and NO env var: the ladder picked native on its
own ("pf-vkdecode auto rung" log line), 525/526 stats windows on
native-vulkan over ~8m46s / 31550 frames, fps 0/59.4/61 with 6
windows <55 incl. startup zeros, bad-signature grep over the whole
log EMPTY, zero TDR events, host service Running after teardown.
H265Planner mirrors the H.264 layer's contract exactly: plan_au -> AuPlan
{ picture, slices with ref lists by stable PicId, DpbUpdate, warnings },
same concealment posture (warnings never abort, in-place reference
substitution preserving ref_idx positions, outputs survive failed AUs,
flush gates on AwaitingIdr, any IRAP resumes). Ported logic: RPS 8.3.2
(short-term AND long-term incl. PocLsbLt/MSB-cycle - the hosts' RFI
recovery rides long-term refs), ref lists 8.3.3/8.3.4, DPB C.5.2.2/C.5.2.3
via the vendored dpb; POC 8.3.1 from the vendored PictureData. Written
fresh: the plan surface, AU walk, envelope gates (multilayer, interlaced,
SCC self-reference, DPB>16, conf-window overflow - checked at EVERY
activation, not just parse), HEVC recovery-point SEI (prefix NALU 39,
se(v) recovery_poc_cnt), VUI colour with E.3.1 inference, and a test-only
HEVC bitstream synthesizer (upstream has none).
Upstream deviations worth naming (all in-code with spec anchors): the
empty-RPS inter slice cannot infinite-loop (upstream bug); RASL behind a
joined CRA refuses BEFORE any state change (PlanError::RaslSkipped - the
WP-2 wiring must map it to skip, not reanchor; module docs carry the
contract note); MaxPicOrderCntLsb reads from the ACTIVATING SPS (upstream
latches at parse - a latent multi-SPS bug); C.5.2.2's exemption is
picture 0 of the BITSTREAM (EobNut), never first-after-EOS.
Vendored parser gained PROVENANCE deviation 7 (report upstream): hostile
slice headers with num_long_term_sps+num_long_term_pics > 16 indexed out
of bounds of SliceHeader's [_;16] arrays - a production panic on exactly
the long-term-reference path, now a parse error.
Port review round 7: 10 findings (3 blocking: the vendor panic, an
EOS-boundary output interleave, an envelope bypass through PPS-only SPS
rebind reaching wrapping crop arithmetic) - 9 fixed with a regression
test each, 1 documented as the WP-2 contract note. Known follow-up: the
h264 AU-tail truncation detector shares h265's dead-arm shape (its arm
also cuts reserved NALU types, so the fix is not identical - deferred).
Tests: 29 h265 planner + 2 HEVC SEI + full test-25fps.h265/bear/bbb clip
walks with real invariants (every stored id output exactly once,
ascending POC per IRAP period). Gates: fmt clean; clippy -D warnings zero
(mac + pf-lxcheck2 incl. pf-client-core/pf-presenter); tests 45+69 mac,
69+121+53 container.
The round-4 residuals, closed after the WP-D hardware verdict:
- VUI colour plumbing (the one silent-wrong): the picture's ACTIVE SPS's
colour signalling (H.273 code points + range, with E.2.1's 'unspecified'
inference where the VUI is silent — the vendored parser's defaults ARE the
inferred values, verified) rides PicturePlan -> DecodedVkFrame ->
NativeVkFrame per frame, never latched: the Windows host switches an HDR
desktop to PQ/BT.2020 IN-BAND while the Welcome still says SDR. Before
this, the native path would have painted PQ washed out, silently.
- Native decode-latency stat: the deliberately-deferred NativeVk arm of the
pump's sampled once-per-stats-window decode measurement now feeds - the
frame's (semaphore, semaphore_value) is the decode-done signal, resolved
through the shipped ledger before a bounded, pure-measurement
vkWaitSemaphores (VkH264Decoder::wait_decoded).
- The renegotiation-teardown window is settled as NO HOLE: rebuild_state now
documents the full safety argument (graveyarded pools stay intact under
presenter holds, tokens route strictly by generation, session objects die
only post-drain with the generation gate INSIDE read_status), and the two
backend comments that wrongly claimed stale pools were 'gone' are fixed.
- VK_KHR_unified_image_layouts stays deferred (fleet drivers lack it).
Adversarial review round 6: 3 minor findings (2 doc fixes applied; the
SPS-replaced-without-PPS-resend divergence stays a documented envelope
assumption - hosts re-send both at every keyframe, and a hardening
PlanWarning could cost real frames on a false positive).
Gates: fmt clean; clippy -D warnings zero (mac + pf-lxcheck2 container,
incl. pf-client-core/pf-presenter); tests 45+30+53 mac, 30+121+53 container.
WP-D parity A/B. gpu_parity (ignored) decodes the conformance vector,
reads every frame back through the presenter's exact contract (wait,
layout round-trip, signal-back, release), crops at the copy so pitch
can never leak, and compares SHA-256s in display order against goldens
from ffmpeg software decode — cross-checked bit-identical between
ffmpeg 8.0.1 (linux) and 8.1.1 (macOS), so the reference is the spec,
not one build. PF_VKD_TEST_READBACK=1 is the one test-only hook (ORs
TRANSFER_SRC into pool usage; production pools stay zero-copy-tight).
Fleet verdict: 250/250 frames bit-identical to libavcodec on RADV
(Mesa 26.0.3, distinct), AMD proprietary Windows (25.10.30.02,
distinct) and NVIDIA Windows (610.88, coincide) — H.264 decode is
exactly specified, and the native path meets the spec on every driver
and both DPB arrangements.
Session-memory and image allocation now honor each binding's
memoryTypeBits with DEVICE_LOCAL preferred, not required: NVIDIA 610.88
legally places a video-session binding in host-visible-only memory and
the hard requirement refused the whole device. The bitstream ring keeps
its hard HOST_VISIBLE|COHERENT need. Smoke test gains
PF_VKD_SMOKE_VENDOR device pinning + attribution and a final-state
print (DPB mode now observed, not inferred).
On-glass matrix after this fix (.173, vendor-pinned): NVIDIA 4090
PASSES in COINCIDE mode — the first end-to-end run of the RESULT_STATUS
query path, ~44 per-frame driver verdicts on the recording pattern that
hangs RADV's VCN — and Adrenalin re-passes in distinct mode unchanged.
With RADV's distinct pass, both DPB arrangements and three of four
desktop drivers are now hardware-validated; Intel remains a clean caps
refusal (no SAMPLED on decode outputs — its rung stays D3D11VA).
Gates: fmt clean, clippy -D warnings zero, 45+27+53 green both
platforms.
WP-D leg 1 (.25 RADV, distinct mode) root causes, both real:
1. Output starvation: the fixed 4-deep ring lost to a stream that keeps
max_dpb_frames+1 = 8 pictures pending. Zero-copy fix (user
requirement, no copies): one picture pool of required_slots +
HOLD_HEADROOM(8) images decoupled from DPB slots — a re-activated
slot binds a fresh free image, so a delivered picture is never a
decode target; the WP-B pin layer became dead and is deleted.
Per-image timeline semaphores carry the AVVkFrame contract: decode
signals value+1, the presenter waits and signals back, later decodes
wait the image's latest value — layout traffic ordered against
reference reads with no copy anywhere.
2. RESULT_STATUS queries HANG RADV's VCN firmware (ring timeout,
DEVICE_LOST): queryResultStatusSupport=false on the decode family.
Queries are now caps-gated; without them poll/wait degrade to
timeline-completion verdicts (FFmpeg parity — and the likely reason
upstream never wired nb_queries). The Ally-X-class detection runs
where drivers advertise the query; .173 probes NVIDIA/Windows-AMD.
Also: slice-only bitstream feeding (the field-proven consumer shape),
graveyarded pool retirement keyed by release tokens + generation,
decode-current-AU-before-status attribution, take_ready drained,
H264-bit gating, teardown short-circuit on disconnected channel.
On-glass: 48 AUs green on .25 holding 4 frames like the real client.
Gates: fmt clean, container clippy -D warnings zero, 27+121+52 green
both platforms.
M2 WP-C. video_vk_native.rs adapts the presenter's VulkanDecodeDevice to
pf-vkdecode (queue lock shared only when the families actually collide —
the one case the 2026-07-09 DEVICE_LOST race proved matters), and the
presenter consumes DecodedImage::NativeVk on its own device: no handle
import, no AVVkFrame co-authoring — wait the timeline, barrier to
sampled, existing crop-aware CSC, barrier back, release after the fence.
Frame lifetime is a token: presented, retired, displaced or dropped
mid-demotion, the guard's drop sends it exactly once; the backend
releases the decoder slot only after the status query resolves, so a
recycled slot can never report a false Failed. Driver-reported decode
failures and plan warnings ride the existing streak/reanchor machinery —
the Ally X corruption class is now a visible error, not a silent frame.
Opt-in only until WP-D's on-glass parity verdict; H.264 sessions only;
failures demote to the existing ladder. Known WP-D items recorded in
code: coincide-mode cross-queue reference overlap, renegotiation
teardown window, VUI colour plumbing.
Gates: fmt clean; container clippy -D warnings zero for pf-client-core +
pf-presenter + pf-vkdecode; 121+53+27 tests green.
M2 WP-B. VkVideoSessionKHR lifecycle with drain-before-destroy on
parameters recreation, DPB pools in both coincide and distinct modes
(caps-derived, usage/flags validated against the driver's format
properties), an aligned bitstream ring, vkCmdDecodeVideoKHR recording
with one-shot RESET re-armed on failed submits, timeline-semaphore
completion, and the per-op RESULT_STATUS query ring — the signal
FFmpeg's hwaccel never reads and the reason this program exists.
Frame lifetime is two-phase by construction: release_frame pins a
delivered frame's slot against reuse, closing the coincide-mode
overwrite the adversarial review round proved (a full DPB handed a
just-returned frame's image back as the same call's decode target).
Nine review findings fixed pre-commit; a counterfactual test pins the
collision. Generation-stamped frames, memory-type misses as errors,
granularity-aligned extents, level gate.
AuPlan now carries its activated SPS/PPS (Rc) so backends never
re-parse. GPU smoke test (ignored) decodes 48 AUs past DPB-full with
releases — the fleet runs it in WP-D.
Gates: fmt clean, clippy -D warnings zero, 45+27+53 tests green on
macOS and the linux/amd64 container.
M2 WP-A (design/client-native-decode.md §3.2). AuPlan -> StdVideo
parameter sets (owned pointer backings), per-AU decode info with slice
start-code offsets, and a PicId->slot map that never evicts on its own.
Deliberate rejections over silent claims: FMO, separate colour planes,
DPBs deeper than 16 frames (unbounded VUI ue(v)) all fail closed.
pf-bitstream API grew what the review proved necessary: RefPic carries
the true top/bottom field order counts (a single poc fabricated
BottomFieldOrderCnt whenever the PPS signals pic-order deltas), the
>16-frame DPB envelope gate, and an MMCO5 rebase warning.
Adversarial review round two: 8 findings fixed pre-commit, including
transactional slot mutation (an error path could permanently desync
the map) and count/pointer coherence on type-1 POC offsets.
Gates: fmt clean, clippy -D warnings zero, 45+26+21 tests green on
macOS and the linux/amd64 container.
Adapted from cros-codecs decoder/stateless/h264.rs (POC 8.2.1, ref lists
8.2.4 incl. modification, sliding-window + MMCO/LTR marking 8.2.5,
frame_num-gap handling 8.2.5.2), minus the fd-coupled backend trait.
H264Planner::plan_au maps one wire AU to picture params, per-slice byte
ranges + ref lists keyed by stable PicIds, and a DPB update; recovery-
point SEI parsing is new code (upstream reads no SEI payloads).
Concealment posture, deliberately different from upstream's aborts:
frame_num gaps, failed RPLM/MMCO and mis-split AUs degrade to warnings
the session turns into recovery asks, gap placeholders substitute
in-place so ref_idx mapping never shifts, and DPB outputs queued during
a failed AU survive to the next plan.
An adversarial review round fixed 11 findings before this commit;
one was an upstream cros-codecs bug our conformance-window test exposed
(display_resolution double-subtracts the crop offset and underflow-
panics — PROVENANCE deviation #6, worth reporting upstream).
Gates: fmt clean, clippy -D warnings zero, 45+24 tests green on macOS
and the linux/amd64 container.
#![forbid(unsafe_code)] on both crates. Upstream's codec module was one
production unsafe away: build_ref_pic_lists turned DPB borrows into
indices via pointer offset_from — same pointer-identity mapping now
expressed as position(ptr::eq) over the <=16-entry DPB (PROVENANCE #5).
Three test-only mem::zeroed() asserts became Default::default(), an
identical value for the all-integer PredWeightTable.
Honest coverage note: build_ref_pic_lists has no callers inside the
vendored subset (its consumer was the non-vendored stateless layer), so
the rewrite is equivalence-by-construction until the DecodePlan layer
exercises it against goldens.
M1 of design/client-native-decode.md. The vendored snapshot (AOSP mirror
main @ 5ff6d693ffae, BSD-3, PROVENANCE.md) is the codec module only —
H.264/H.265/AV1/VP9 parsers, DPBs and their test vectors, which now run
as 45 conformance tests in our CI. pf-bitstream sits where upstream's
Linux-only decoder::stateless half would and starts with vendor-pinning
smoke tests: a re-sync that shifts parser behavior trips in-tree, not in
a decode session.
Fixture-corpus enabler for the native-decode program (M0,
design/client-native-decode.md): every AU exactly as the pump hands it
to decode_frame — the raw concatenation plus a sidecar .idx carrying
the AU boundaries and wire flags a byte stream cannot. Best-effort by
design: any I/O error disables the capture, never the stream.
A minor bump: 98 commits since v0.24.0. The headline is DualSense pad audio
(PR #23) — a wired DualSense playing a game's voice-coil haptics and its own
speaker, streamed from the host, on Android and the desktop session client
against a Windows host with Steam's driver present. Behind it: the haptics
sweep's twelve milestones closing more than twenty controller faults across
every client and both hosts; the audio quality/latency work (256 kbps stereo,
the Steam Streaming Microphone endpoint root cause, and the de-jitter ratchet
that left audio permanently behind the picture); and MTU resilience plus
mid-session shard renegotiation, which turns the silent all-black stream on a
sub-1330-byte path into a diagnosed warning that heals itself. Plus the Decky
plugin reduced to a launcher, system-button routing with hold-Select, gamepad-UI
profiles on all three UIs, `discover`/`launch --request-access` in the CLI, and
the Sunshine false-conflict and crashed-host display-restore fixes.
The canary base is already 0.25 — scripts/ci/pf-version.sh derives it as one
minor ahead of the latest stable tag — so this is the version canary has been
publishing against all along.
Wire protocol stays at 2: every addition this cycle is optional or
capability-gated (an optional trailing max_shard_payload on Hello, the 0x08/0x09
renegotiation pair, the 0xD1 pad-audio plane, the 0xD2 redundant desktop-audio
plane, MAX_DATAGRAM_BYTES 2048 -> 9216). C ABI moves 14 -> 16 in two steps: 15
retroactively declares the floor that guarantees the rumble policy engine's C
surface (which shipped while the constant still read 7), and 16 adds the
pad-audio surface and mirrors its two capability bits. Four new capability bits
land in the client/host bytes (audio redundancy 0x04/0x20, pad audio 0x08/0x40);
the video-caps byte was NOT touched and stays full from 0.23.0, so the standing
"next video cap needs a second byte and an ABI bump" note still holds. host_caps
is now down to its last free bit (0x80). Virtual-display driver protocol 6 and
the Windows gamepad channel 3 are untouched — pf-driver-proto is byte-for-byte
identical to v0.24.0. The generated header is in sync (ABI 16, both cap mirrors).
Breaking for C embedders: 149 unprefixed macros are now PUNKTFUNK_-prefixed
(139 #defines renamed in the checked-in header). Mechanical to fix, and it
cannot break silently — the old spellings cease to exist, so it is always an
undeclared-identifier error rather than the wrong value a colliding #define
used to produce.
Lock touched for the 32 workspace members only, via `cargo update --workspace`:
diff against origin/main is versions-only, 32 insertions and 32 deletions. Unlike
the last cut there is no third-party crate sitting on the outgoing version to
trip the count — `wasapi` is at 0.23.0 and was never a candidate. `cargo metadata
--locked` resolves (35 members; fec-rs, pf-driver-proto and usbip-sim keep their
own versions by design). `cargo fmt --all --check` clean in both the main and
packaging/windows/drivers workspaces. Doc lazy-continuation scanner: 0 hits over
521 files — that regex is the exact defect that made the first v0.23.0 tag go red
on Windows clippy, and no Windows leg runs on a main push, so main being green
proves nothing about the tag fan-out.
api/openapi.json is deliberately left at 0.23.0: it tracks API edits and lags,
as in every prior cut. It is now two releases behind and worth a look.
Notes at docs/releases/v0.25.0.md, per docs/releases/README.md — authored with
the bump so CI's ensure_release seeds the release body at tag creation. Body
voice checked programmatically: 0 internal-vocabulary hits above `## Under the
hood`. Play's "What's new" at docs/releases/whatsnew/v0.25.0.txt (494/500 chars),
verified by running android.yml's gate logic verbatim against it, including the
byte-identical-to-another-release check.
2026-08-05 00:05:40 +02:00
594 changed files with 152476 additions and 18790 deletions
# `punktfunk-canary` pacman repo as X.Y.Z-0.<run#> (sorts below the eventual X.Y.Z-1),
# tags to `punktfunk` — separate repos, so neither channel can shadow the other.
tags:['v*']
# REBUILDING A PUBLISHED RELEASE, because on a rolling distro the ground moves under one.
# Arch went FFmpeg 8 -> 9 (every libav soname +1) four minutes before v0.25.0 was tagged, so
# the release's punktfunk-host was linked in a builder image that still had 8 and shipped
# `libavcodec.so=62-64`. No up-to-date Arch box can satisfy that — and pacman prepares the
# whole transaction at once, so it did not merely block our package, it blocked those users'
# entire `pacman -Syu`. The repair is a rebuild of the SAME upstream version at a HIGHER
# pkgrel; nothing else reaches a box that already has the broken build recorded in its db.
# The workflow file at the tag can never carry inputs added after it was tagged, so dispatch
# this from `main`: it checks the tag's SOURCE out, publishes to the STABLE repo, and
# replaces the release-page assets. Same lever for any future "the distro moved" rebuild.
workflow_dispatch:
inputs:
release_tag:
description:'Rebuild this published release (e.g. v0.25.0) into the stable `punktfunk` repo. Empty = ordinary canary build of the dispatched ref.'
required:false
default:''
pkgrel:
description:'pkgrel for that rebuild — MUST be above the published one (2, 3, …); a same-pkgrel republish is invisible to pacman. Ignored without release_tag.'
required:false
default:'2'
env:
REGISTRY:git.unom.io
@@ -94,7 +113,52 @@ jobs:
}
bun --version
# THE BUILDER'S FFmpeg IS PART OF THE PACKAGE CONTRACT, not merely a build detail.
# packaging/arch/PKGBUILD binds punktfunk-host to the exact libav sonames it linked
# (`libavcodec.so=63-64` …), so a builder one FFmpeg major behind Arch emits a package
# that NOBODY can install — and takes the user's whole `pacman -Syu` down with it, since
# pacman prepares the transaction as a unit. That is exactly how v0.25.0 shipped: PR #108
# re-keyed this image for FFmpeg 9, the release tag fired four minutes later, and the job
# still got the FFmpeg-8 `:latest`. The image is a cache and is allowed to lag — but never
# on this one axis. So heal it in-job and shout, instead of building a dead package.
# (Runs BEFORE checkout: a stale image should be repaired before anything depends on it.)
- name:FFmpeg soname parity with today's Arch (heals a stale builder image)
run:|
export LC_ALL=C # `Provides` is a localized field name
# Piped (never a TTY here) pacman prints each field on ONE line, unwrapped.
| python3 -c "import json,sys;k=set(sys.argv[1].split());k|={n+'.sha256' for n in k};print('\n'.join('%s %s'%(a['id'],a['name']) for a in json.load(sys.stdin) if a.get('name','').endswith(('.pkg.tar.zst','.pkg.tar.zst.sha256')) and a['name'] not in k))" "$KEEP" \
volume/routing bytes, change-only and value-deduped. Older clients drop it as an unknown kind.
- **Arrival flags** — bits 8 (haptics) and 9 (speaker), sent only toward a `HOST_CAP_PAD_AUDIO` host.
- **Adaptive-trigger effects are length-bounded** on encode and decode against one shared constant;
the header emits `uint8_t effect[PUNKTFUNK_HID_EFFECT_MAX]` in place of a literal `11` (same value,
so the struct layout is byte-identical). A zero-length effect body is now rejected rather than
decoding as an empty — that is, a *release* — effect.
- Out-of-range pad indices are dropped before **either** rumble consumer sees them. The reorder gate
bounds-checked and the legacy queue did not, so an embedder draining it could be handed an index it
would use to subscript its own array. The client also clamps the host's rumble lease receive-side
at 5 s, where the ceiling had been sender-side only.
### Host environment variables
| Variable | Default | Notes |
|---|---|---|
| `PUNKTFUNK_AUDIO_QUALITY` | `high` | `low`/`standard`/`high`; `high` = stereo 256 kbps. `standard` reproduces the pre-0.25 encoder exactly for an A/B. A typo warns once rather than silently downgrading. |
| `PUNKTFUNK_AUDIO_REDUNDANCY` | unset = automatic | on when the client supports it and the budget allows |
| `PUNKTFUNK_LIBRARY_ART_ROOTS` | platform default | art-serving roots; POSIX now defaults to `$HOME` |
| `PUNKTFUNK_DECODER` | client | **values changed**: `native-vulkan` · `native-vaapi` (Linux) · `native-d3d11va` (Windows) · `software`. Legacy `vulkan`/`vaapi`/`d3d11va` still accepted and migrated. Now **trimmed** — a trailing space used to fall through to `auto` silently. |
"description":"Every installed-store title (Steam, read from the host's local files — no Steam API key)\nmerged with the user's custom entries, sorted by title. Artwork fields are URLs the client\nfetches directly (the public Steam CDN for Steam titles). `?provider=` narrows to the\nentries a given external provider owns; `?platform=` to one platform (case-insensitive —\ninstalled-store titles are `PC`, custom/provider entries carry whatever was authored).",
"description":"Every installed-store title (Steam, read from the host's local files — no Steam API key)\nmerged with the user's custom entries, sorted by title. Artwork fields are URLs the client\nfetches directly (the public Steam CDN for Steam titles). `?provider=` narrows to the\nentries a given external provider owns; `?platform=` to one platform (case-insensitive —\ninstalled-store titles are `PC`, custom/provider entries carry whatever was authored).\n\n**The operator's own lane additionally sees the titles they have HIDDEN**, each carrying\n`hidden: true`; every other lane gets them filtered out upstream and cannot tell they exist. The\nconsole needs them to offer \"un-hide\", and it is the only surface that does.",
"operationId":"getLibrary",
"parameters":[
{
@@ -1021,13 +1021,13 @@
],
"responses":{
"200":{
"description":"Unified library across all stores",
"description":"Unified library across all stores (the operator's lane also gets hidden entries, flagged)",
"content":{
"application/json":{
"schema":{
"type":"array",
"items":{
"$ref":"#/components/schemas/GameEntry"
"$ref":"#/components/schemas/OperatorGameEntry"
}
}
}
@@ -1301,6 +1301,79 @@
}
}
},
"/api/v1/library/hidden/{id}":{
"put":{
"tags":[
"library"
],
"summary":"Hide or un-hide one library title",
"description":"Curation, not access control: a hidden title disappears from every play surface — the console\ngrid on a client, native clients, the GameStream app list, and launch resolution — while nothing\nis deleted and un-hiding restores it immediately. The operator's own console still lists it\n(flagged `hidden`) so it can be brought back.\n\nKeyed by the entry's stable `<store>:<external_id>` id, which survives re-scans and reconciles by\nconstruction (D2). The id is **not** validated against the current library on purpose: a title\ncan be legitimately absent at this moment (launcher closed, plugin mid-sync, drive unmounted),\nand refusing the operator's choice in that window would be worse than storing an id that\ncurrently matches nothing. Emits `library.changed` (source = the store) only on a real change.",
"operationId":"setLibraryEntryHidden",
"parameters":[
{
"name":"id",
"in":"path",
"description":"The library entry id (e.g. `steam:70`)",
"required":true,
"schema":{
"type":"string"
}
}
],
"requestBody":{
"content":{
"application/json":{
"schema":{
"$ref":"#/components/schemas/HiddenToggle"
}
}
},
"required":true
},
"responses":{
"200":{
"description":"Stored; the entry's visibility after the call",
"content":{
"application/json":{
"schema":{
"$ref":"#/components/schemas/HiddenState"
}
}
}
},
"400":{
"description":"Empty entry id",
"content":{
"application/json":{
"schema":{
"$ref":"#/components/schemas/ApiError"
}
}
}
},
"401":{
"description":"Missing or invalid bearer token",
"content":{
"application/json":{
"schema":{
"$ref":"#/components/schemas/ApiError"
}
}
}
},
"500":{
"description":"Could not persist the settings",
"content":{
"application/json":{
"schema":{
"$ref":"#/components/schemas/ApiError"
}
}
}
}
}
}
},
"/api/v1/library/provider/{provider}":{
"put":{
"tags":[
@@ -4045,6 +4118,51 @@
}
}
},
"AudioWiring":{
"type":"object",
"description":"The Windows host's audio wiring verdict — which endpoint carries each role. The names are\nthe endpoints' friendly names as the Sound settings show them (on current hosts the minted\n\"Punktfunk\" instances of Steam's streaming drivers).",
"required":[
"readiness",
"mic_withheld",
"last_resort"
],
"properties":{
"last_resort":{
"type":"boolean",
"description":"The loopback is the known-degraded last resort — desktop audio may be silent until the\nendpoint set changes."
},
"loopback":{
"type":[
"string",
"null"
],
"description":"Friendly name of the desktop-audio loopback source; absent = desktop audio unavailable."
},
"mic":{
"type":[
"string",
"null"
],
"description":"Friendly name of the virtual-mic write target; absent = mic passthrough unavailable."
},
"mic_withheld":{
"type":"boolean",
"description":"The mic was WITHHELD so game audio could keep the only working sink — mic passthrough\nneeds Steam installed (the host mints its own microphone) or a virtual cable."
},
"narrowing":{
"type":[
"string",
"null"
],
"description":"Why the chosen loopback endpoint NARROWS the desktop mix (rate/channels), when it does."
},
"readiness":{
"type":"string",
"description":"`full` | `audio_only` | `mic_only` | `none` — whether desktop audio and mic passthrough\neach have an endpoint at all.",
"example":"full"
}
}
},
"AvailableCompositor":{
"type":"object",
"description":"A compositor backend the host can drive a virtual output on, and whether it's usable now.",
"description":"Request body for `setLibraryEntryHidden`.",
"required":[
"hidden"
],
"properties":{
"hidden":{
"type":"boolean",
"description":"Whether this title should be hidden from every play surface."
}
}
},
"HookEntry":{
"type":"object",
"description":"One hook: fire `run` and/or `webhook` when an event matching `on` (+ `filter`) occurs.",
@@ -6294,6 +6443,23 @@
}
}
},
"OperatorGameEntry":{
"allOf":[
{
"$ref":"#/components/schemas/GameEntry"
},
{
"type":"object",
"properties":{
"hidden":{
"type":"boolean",
"description":"The operator hid this title ([`set_entry_hidden`]) — omitted when false, so the shape only\ngrows for entries that actually are hidden."
}
}
}
],
"description":"A library entry plus the operator's own view of it — today, whether they hid it.\n\nA separate type rather than a field on [`GameEntry`] for two reasons. It keeps the visibility\nanswer out of the providers entirely: a store parser has no opinion on what the operator hid, and\nadding `hidden: false` to all eight construction sites would imply it does. More importantly it\nmakes the lane rule a TYPE guarantee instead of a discipline — `GET /library` answers\n`Vec<GameEntry>` on every lane but the operator's, so a hidden entry cannot leak to a paired\nclient by someone forgetting a filter; there is no field there to leak.\n\n`flatten` keeps the wire shape identical to a plain entry with one extra key, so the console\nparses one model either way."
"description":"Number of live streaming sessions across BOTH planes (GameStream + native punktfunk/1). The\nnative server admits concurrent sessions, so this can exceed 1; `session`/`stream` below\ndescribe a single representative session for the detail card.",
"minimum":0
},
"audio":{
"oneOf":[
{
"type":"null"
},
{
"$ref":"#/components/schemas/AudioWiring",
"description":"The audio wiring verdict (Windows hosts; absent on other platforms and before the first\nwiring pass). Present even while idle — the wiring exists for the host's lifetime."
}
]
},
"audio_streaming":{
"type":"boolean",
"description":"True while the audio stream thread is running."
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.