diff --git a/quest/m1/vr-demo-try.md b/quest/m1/vr-demo-try.md index b059e90ce26..352b62cf59c 100644 --- a/quest/m1/vr-demo-try.md +++ b/quest/m1/vr-demo-try.md @@ -2,8 +2,8 @@ ## Goal -The maintainer has played `demo/vr` in a WebXR headset, and once it plays -well there, it is hosted beside the web demo at moq.dev with a looping VR180 +Someone has played `demo/vr` in a WebXR headset, and once it plays well +there, it is hosted beside the web demo at moq.dev with a looping VR180 broadcast to watch. ## Plan @@ -22,6 +22,10 @@ headset's localhost, and make the relay reachable too: `localhost:4443` on the headset is the headset itself, so forward that port as well or point the page at the relay with `?url=`. +Decided 2026-11-15 (maintainer, /quest-complete): ✅ ask the contributor who +started the demo in [moq-dev/moq#6373](https://github.com/moq-dev/moq/pull/6373) +to try it in a headset, rather than the maintainer. + ## Waiting -- Maintainer: try it in a headset (an agent has none) and say whether it plays well enough to host. Check by this quest's Plan recording the result. +- nuts-rice: try `demo/vr` in a WebXR headset (an agent has none) and say whether the per-eye render, recentring, and ambisonic audio play well enough to host, as asked in [moq-dev/moq#8781](https://github.com/moq-dev/moq/pull/8781#issuecomment-6901206041) on 2026-11-15. Check by reading that thread for a reply; this quest's Plan records the headset result. diff --git a/quest/m2/websocket-accept-driver.md b/quest/m2/websocket-accept-driver.md new file mode 100644 index 00000000000..079ac44fd55 --- /dev/null +++ b/quest/m2/websocket-accept-driver.md @@ -0,0 +1,31 @@ +# [S] moq-tokio's WebSocket accept hands back the qmux driver + +## Goal + +`moq_tokio::websocket::accept` returns the qmux driver to its caller instead +of spawning it on the tokio runtime, so a caller such as an io_uring worker +runs the driver on its own task set. Every existing caller (moq-relay's +`websocket.rs` and moq-tokio's own server tests) spawns the driver itself and +behaves as today. + +## Plan + +Decided 2026-11-15 (maintainer, /quest-complete): ✅ reshape +`websocket::accept` itself to return the session and its driver, as its own +m2 quest that [relay stream listeners](/quest/m9/stream-transports/relay.md) +requires (rejected: landing it with the rust2 migration; an additive sibling +beside `accept`). + +Today `accept` builds the qmux session, hands it to `spawn_qmux` +(`rs/moq-tokio/src/lib.rs`), which spawns the `Driver` on the current +runtime, and then reads SETUP into a `server::Request` +(`rs/moq-tokio/src/websocket.rs`). The SETUP read needs the driver running, +so the driver cannot come back only at the end: return it once QMux is +established, beside whatever finishes the SETUP read and admission, and keep +the accept deadline covering both. + +This breaks moq-tokio's published API, so the PR's title takes `!` and its +Impact section carries the migration note: a caller that used the old +`accept` spawns the returned driver itself. + +Public API: breaking, `moq_tokio::websocket::accept`'s return type. Wire: none. diff --git a/quest/m3/capture-linux.md b/quest/m3/capture-linux.md index 0578b34a721..94c0d5a0934 100644 --- a/quest/m3/capture-linux.md +++ b/quest/m3/capture-linux.md @@ -2,8 +2,9 @@ ## Goal -Window capture and system audio work on a Wayland desktop, and app capture has -a stated answer rather than a runtime error. +Window capture and system audio work on a Wayland desktop. App capture stays +macOS-only (decided below), which `Source::App` already reports as +`Unsupported` here. ## Plan @@ -31,9 +32,15 @@ unimplemented, because xdg-desktop-portal owns source selection: - **System audio** needs a PipeWire monitor node. `moq_audio::capture::Source::System` exists but is macOS-only and returns `Unsupported` elsewhere. This is independent of the screencast portal. -- **Applications** have no portal concept behind them, so decide explicitly - whether Linux offers app capture rather than leaving a variant that errors at - runtime. +- **Applications** have no portal concept behind them. Decided 2026-11-15 + (maintainer, /quest-complete, with + [Windows capture parity](/quest/m3/capture-windows.md)): ✅ (A) Linux offers + window capture only: `Source::App` stays macOS-only and fails with its + reason here, and a listing uses apps only to group windows + (rejected: (B) build app capture by composing an app's windows). No + consumer asks for it, and it matches the + [capture listing](/quest/m3/capture-list.md)'s rule that a listed app must + not hand out a source that fails at open. - **Per-app audio** (decided 2026-10-26, maintainer: yes) captures one application's output from its PipeWire stream node, so a creator can stream a game without their music. Apps are listed from the nodes present, so only @@ -43,10 +50,6 @@ unimplemented, because xdg-desktop-portal owns source selection: Windows process tree and the macOS application. The Windows and macOS halves share the same `moq_audio::capture` source. -## Waiting - -- Maintainer: whether app capture exists off macOS, here and in [Windows capture parity](/quest/m3/capture-windows.md). (A) Linux and Windows offer window capture only: `Source::App` stays macOS-only and fails with its reason elsewhere, and a listing off macOS uses apps only to group windows. (B) Build app capture by composing an app's windows. I recommend A, since no consumer asks for B and it matches the capture listing's rule that a listed app must not hand out a source that fails at open. - ## Related - [Windows capture parity](/quest/m3/capture-windows.md) - the same gaps, diff --git a/quest/m3/capture-windows.md b/quest/m3/capture-windows.md index 1e9d6ffe138..38ce9440170 100644 --- a/quest/m3/capture-windows.md +++ b/quest/m3/capture-windows.md @@ -2,7 +2,8 @@ ## Goal -App capture and system audio work on Windows. Display and window capture, +System audio works on Windows. App capture stays macOS-only (decided below), +which `Source::App` already reports as `Unsupported` here. Display and window capture, including the cursor, already run on the Windows.Graphics.Capture backend (`rs/moq-video/src/capture/wgc/native.rs`), which [its own quest](/quest/m3/capture-wgc.md) verifies on real hardware. @@ -11,9 +12,14 @@ including the cursor, already run on the Windows.Graphics.Capture backend - **Applications** (every window of a process, including ones that open later) are still `Unsupported` outside macOS, where `SCShareableContent` gives it - almost free. Windows has no direct equivalent, so this needs per-window - composition or an explicit decision that Windows offers window capture only. - Decide that rather than leaving a variant that errors at runtime. + almost free. Decided 2026-11-15 (maintainer, /quest-complete, with + [Linux capture parity](/quest/m3/capture-linux.md)): ✅ (A) Windows offers + window capture only: `Source::App` stays macOS-only and fails with its + reason here, and a listing uses apps only to group windows + (rejected: (B) build app capture by composing a process tree's windows). No + consumer asks for it, and it matches the + [capture listing](/quest/m3/capture-list.md)'s rule that a listed app must + not hand out a source that fails at open. - **System audio** needs WASAPI loopback. `moq_audio::capture::Source::System` exists but is macOS-only and returns `Unsupported` elsewhere. Unlike macOS this does not go through the screen-capture API, so it is an independent path @@ -26,14 +32,6 @@ including the cursor, already run on the Windows.Graphics.Capture backend setup and refuse per-app audio with the reason where activation fails, rather than gating on a build number. -WGC does not answer app capture. It captures one monitor or one window per -item and has no per-process source, so the per-window-composition or -window-only decision still stands now that the backend has landed. - -## Waiting - -- Maintainer: whether app capture exists off macOS, here and in [Linux capture parity](/quest/m3/capture-linux.md). (A) Windows and Linux offer window capture only: `Source::App` stays macOS-only and fails with its reason elsewhere, and a listing off macOS uses apps only to group windows. (B) Build app capture by composing a process tree's windows. I recommend A, since no consumer asks for B and it matches the capture listing's rule that a listed app must not hand out a source that fails at open. - ## Related - [Windows.Graphics.Capture](/quest/m3/capture-wgc.md) - display and window capture with the cursor, which this quest no longer owns diff --git a/quest/m3/cli-packaging.md b/quest/m3/cli-packaging.md index 345e722e53a..43024d109d8 100644 --- a/quest/m3/cli-packaging.md +++ b/quest/m3/cli-packaging.md @@ -68,9 +68,14 @@ and `moq-audio/capture` explicitly, and `full` enables `capture`. Renamed 2026-11-07 (maintainer, /quest-plan from the API sweep of open PRs and remaining quests): the moq-cli feature that enables everything is named `full`, not `release`, so this quest and its followers say `full`. -## Waiting - -- Maintainer: what ships while cpal has no release that loads libasound at runtime (RustAudio/cpal#1421 is open; the last release is 0.18.2). (A) Wait: this quest stays blocked on [ALSA without a load-time link](/quest/m3/capture-alsa-link.md) until the cpal release ships. (B) Split: ship `full` now on macOS and Windows, which link no system audio library, and leave only Linux waiting on the cpal release. (C) Linux release binaries take libasound as a load-time requirement. (D) Fork cpal, alsa, and alsa-sys. I recommend B. +Decided 2026-11-15 (maintainer, /quest-complete): ✅ (A) wait for the cpal +release that loads libasound at runtime (RustAudio/cpal#1421 is open; the last +release is 0.18.2), so this stays blocked on +[ALSA without a load-time link](/quest/m3/capture-alsa-link.md), and ✅ (D) +fork cpal, alsa, and alsa-sys only once the wait becomes a blocker (rejected: +(B) ship `full` on macOS and Windows first and leave Linux waiting; (C) Linux +release binaries take libasound as a load-time requirement). In the +maintainer's words: "wait, then fork cpal when it's a blocker". ## Required diff --git a/quest/m3/cpal-alsa-runtime.md b/quest/m3/cpal-alsa-runtime.md index 0fcc65f7061..6d00d912930 100644 --- a/quest/m3/cpal-alsa-runtime.md +++ b/quest/m3/cpal-alsa-runtime.md @@ -64,8 +64,10 @@ fall-through goal, so prefer it. Fallback if upstream declines: fork `cpal`, `alsa`, and `alsa-sys` as published crates, as moq-v4l was. It is the only in-tree route that reaches -`cargo install moq-cli`, at the cost of carrying three forks for one library; -the maintainer decides if it comes to that. +`cargo install moq-cli`, at the cost of carrying three forks for one library. +Decided 2026-11-15 (maintainer, /quest-complete, on +[cli-packaging](/quest/m3/cli-packaging.md)): keep waiting, and take the +fork only once the wait becomes a blocker. Rejected in-tree routes, for whoever retries this: a `[patch.crates-io]` fork of `alsa-sys` does not reach `cargo install moq-cli` or crates.io consumers diff --git a/quest/m3/ffi-exporter.md b/quest/m3/ffi-exporter.md index 149b93a8630..b7aa4a4bf51 100644 --- a/quest/m3/ffi-exporter.md +++ b/quest/m3/ffi-exporter.md @@ -109,6 +109,11 @@ dependency, so crates.io moq-ffi still builds (see the with an error, which the fork still leaks (moq-ffi's errors hold no handles today). +Deferred 2026-11-15 (maintainer, /quest-complete): the `Waiting` decision on +the generated binding renames stays open, kept out of that run's m0 to m2 +scope, since the binding API changes anyway in the rust2 migration expected +around 2026-12-15. + Public API: moq-ffi's dependency names change, generated C type names become CamelCase, and the generated record `MoqAudioEncoderFinished` becomes `MoqAudioEncoderFinish` in every language; wrapper APIs do not change except diff --git a/quest/m3/gpu-ci.md b/quest/m3/gpu-ci.md index e2e44ca64b9..8d7eccf9617 100644 --- a/quest/m3/gpu-ci.md +++ b/quest/m3/gpu-ci.md @@ -47,28 +47,24 @@ skipping inside the Nix shell. NVENC and CUDA external-image import (the verification #4975 deferred here). - Merge queue: a `merge_group` job in `.github/workflows/check.yml` (or `video.yml`) runs `nix develop --command just rs nvidia`, a recipe and not - a script path, like every other workflow step, on a self-hosted runner VM + a script path, like every other workflow step, on a self-hosted runner labeled for the NVIDIA host. Scope it like the other queue jobs: run when the queue's diff touches moq-video, moq-nvenc, or what they depend on. Read-only `permissions`, never `pull_request`. - Decided 2026-10-20 (maintainer): GPU tests gate in the merge queue on the - NVIDIA card, because they gate correctness, not trends. The runner is a - persistent VM from the `github-runner-vm` module (#6040, - `nix/runner/README.md`). That module has no GPU passthrough today: add - VFIO passthrough of the NVIDIA card to the guest, with the driver in the - guest, and keep the runner registered only to this repository with - approval required for every external contributor's workflow, as the - README requires. -- Open since 2026-10-25: the [CI epic](/quest/m0/ci/README.md) dropped the - VM for the CI host in favor of kernel sandboxing, which conflicts with the - VFIO passthrough above. See `## Waiting`. + NVIDIA card, because they gate correctness, not trends. Keep the runner + registered only to this repository with approval required for every + external contributor's workflow. +- Decided 2026-11-15 (maintainer, /quest-complete): ✅ (a) a sandboxed runner + container on the CI host, like `nixos`'s existing runner containers, with + the NVIDIA device nodes and the host driver's libraries bound in, as + [GPU runners](/quest/m9/video-quality/gpu-runners.md) plans for the quality + job (rejected: (b) a VFIO-passthrough VM from `github-runner-vm`, which + conflicts with the [CI epic](/quest/m0/ci/README.md)'s choice of kernel + sandboxing over a VM). Public API: none. Wire: none. -## Waiting - -- Maintainer: pick the GPU runner's shape, since the 2026-10-20 VM decision conflicts with the CI epic's later no-VM decision. Options: (a) a sandboxed runner container on the CI host with the NVIDIA device nodes and the host driver's libraries bound in, as [GPU runners](/quest/m9/video-quality/gpu-runners.md) already plans for the quality job; (b) a VFIO-passthrough VM from `github-runner-vm`. Recommended: (a), then drop the VFIO text above. - ## Related - [Video hardware validation](/quest/m5/video-hardware.md) - hardware paths nothing runs yet diff --git a/quest/m3/lite-tail-drop-count.md b/quest/m3/lite-tail-drop-count.md index 699db91d210..4b24822093c 100644 --- a/quest/m3/lite-tail-drop-count.md +++ b/quest/m3/lite-tail-drop-count.md @@ -13,11 +13,7 @@ Checked 2026-11-12 at a86433043: the counting half already exists. `Tail::accoun - Count per version. lite-01..04 have no drop message, so they keep today's behaviour. - On lite-05/06, account for holes left by lost datagrams. Lite-07 stream tracks carry no datagrams, which have their own track kind. - Account for a stream reset before its header, so no gap is left only to the grace. Check whether the publisher already reports such a group in a SUBSCRIBE_DROP; if it needs a new signal, that is a lite-07-wip draft change, and the Wire line below changes with it. -- Guard against a peer that never accounts for a group as the `Waiting` decision settles. Don't add a new grace. +- Guard against a peer that never accounts for a group. Decided 2026-11-15 (maintainer, /quest-complete): ✅ (c) wait out the subscription's max delay where one is set, and ✅ (a) end the tail with an error on lite-07 when no max delay is set, so the 1 s constant goes away without a buggy peer hanging the tail forever (rejected: (a) alone, an error even where a max delay is set; (b) keep waiting until the session closes). ✅ A lite-05/06 tail with no max delay ends with the same error (rejected: keeping the 1 s constant there). Don't add a new grace. - The IETF tail keeps its grace until the moq-transport gap-signal extension (planned in [moq-dev/moq#7425](https://github.com/moq-dev/moq/pull/7425)) lands, since moq-transport can't signal gaps. Public API: none. Wire: none. - -## Waiting - -- Maintainer: decide what a lite-05+ tail does when the peer never accounts for a group. Options: (a) end the tail with an error; (b) keep waiting until the session closes; (c) wait out the subscription's max delay where one is set. Recommended: (c), and (a) on lite-07 when no max delay is set, so the 1 s constant goes away without a buggy peer hanging the tail forever. diff --git a/quest/m3/webrtc-dial-policy.md b/quest/m3/webrtc-dial-policy.md index f992615f2f6..ab3415bdefd 100644 --- a/quest/m3/webrtc-dial-policy.md +++ b/quest/m3/webrtc-dial-policy.md @@ -45,16 +45,17 @@ Today: Guidance: - One policy type serves both crates, so a host states its rule once for - every destination kind. Propose its shape and its home; a value the host - builds is preferred over a callback, per the root `AGENTS.md`. + every destination kind. Decided 2026-11-15 (maintainer, /quest-complete): + ✅ (a) a plain value type with no I/O in moq-net, which moq-webrtc and + moq-rtmp both already depend on, each crate checking it at connect time. + A host passes it through moq-webrtc's `#[non_exhaustive]` `client::Config` + and moq-rtmp's `#[non_exhaustive]` `Connect`, so neither needs a second + break (rejected: (b) moq-tokio, beside its own dial dispatch, though + moq-rtmp uses moq-tokio only as a dev-dependency; (c) a crate of its own). + A value the host builds, not a callback, per the root `AGENTS.md`. - A refusal is terminal for that URL, and its error is distinct from a failed connection, so a caller that redials on failure does not retry it. - Check at connect time on the resolved address. Checking the URL and then letting the HTTP client resolve again is the DNS rebinding hole. - moq-srt's client takes a `SocketAddr` like moq-rtmp does today. It is out of scope until it dials a URL. - -## Waiting - -- Maintainer: where the one dial-policy type lives. (a) A plain value type with no I/O in moq-net, which moq-webrtc and moq-rtmp both already depend on, each crate checking it at connect time; (b) moq-tokio, beside its own dial dispatch, though moq-rtmp uses moq-tokio only as a dev-dependency today; (c) a crate of its own. Recommended: (a), with the moq-webrtc half added through its `#[non_exhaustive]` `client::Config`, and moq-rtmp's through its `#[non_exhaustive]` `Connect`, so neither needs a second break. - diff --git a/quest/m4/announce-interest-memory.md b/quest/m4/announce-interest-memory.md index be3cc0dc964..25fe256884d 100644 --- a/quest/m4/announce-interest-memory.md +++ b/quest/m4/announce-interest-memory.md @@ -1,13 +1,14 @@ -# [M] Announce interest streams share one route table +# [S] Cap the legacy versions' announce interest streams ## Goal -An announce-interest stream (lite ANNOUNCE_PLEASE or moq-transport -SUBSCRIBE_NAMESPACE) costs memory and CPU proportional to what it has in -flight, not to every route under its prefix. A subscriber opening thousands -of root-prefix streams on a relay holding N routes no longer makes it hold -thousands of times N entries, and a benchmark swept over streams and routes -shows per-stream cost flat in N. +A peer can open only as many announce-interest streams (lite-05/06 +ANNOUNCE_PLEASE or moq-transport SUBSCRIBE_NAMESPACE) as a new +`session::Limits` field allows, so a subscriber opening thousands of +root-prefix streams on a relay holding N routes can no longer make it hold +thousands of times N entries. A stream past the cap is refused, and a test +shows the refusal on each of those versions while a stream under the cap is +served as today. ## Plan @@ -30,22 +31,18 @@ Once [one route stream](/quest/m2/cluster-routes/route-stream.md) lands, lite-07 holds one interest stream per side, so what remains exposed is lite-05/06 and moq-transport. -Bound it by design, not with a cap: cursors walk a shared, refcounted table -snapshot lazily, materializing only what is in flight and gated on stream -write credit; lite-05/06 encodes its initial set incrementally under -`poll_flush`, as lite-07 does. The IETF SUBSCRIBE_NAMESPACE handler follows -the same pattern and gets the same fix. - -Public API: none. Wire: none. - -## Waiting - -- Maintainer: how to bound the legacy versions' interest streams: the shared - snapshot redesign above, a `session::Limits` cap on outgoing interest - streams (which the Plan rules out today), or deleting this once lite-07 is - the oldest version served. Recommended: the cap for lite-05/06 and - moq-transport now, then delete, since redesigning `model/origin.rs` for - versions being retired costs more than it saves. +Decided 2026-11-15 (maintainer, /quest-complete): ✅ a `session::Limits` cap +on interest streams for lite-05/06 and moq-transport now, counted like +`announces` with a `Slots` acquire where each stream starts, and deleted with +this concern once lite-07 is the oldest version served (rejected: a shared, +refcounted table snapshot that cursors walk lazily, since redesigning +`model/origin.rs` for versions being retired costs more than it saves; and +deleting this later with no cap meanwhile). `Limits` is `#[non_exhaustive]`, +so the field is additive. Report its peak in the stats sessions rows, as +`announces` does, and mirror the field in `js/net` if it serves these +streams. + +Public API: additive, a new `session::Limits` field. Wire: none. ## Related diff --git a/quest/m4/binding-export-coverage.md b/quest/m4/binding-export-coverage.md index 7191cf9a4a5..9d75263cc8c 100644 --- a/quest/m4/binding-export-coverage.md +++ b/quest/m4/binding-export-coverage.md @@ -11,16 +11,16 @@ record without an alias still passes. Parity depends on review. ## Plan Found by the 2026-11-07 API shape audit (bindings). Generate the list of -FFI items from UniFFI's metadata once, and have each wrapper's test check -it against the wrapper's public types, with an explicit allowlist for items -a wrapper deliberately hides. Kotlin's -fixed list moves to the shared one. When [the exporter](/quest/m3/ffi-exporter.md) replaces UniFFI, -the list comes from it. +FFI items from [the exporter](/quest/m3/ffi-exporter.md), and have each +wrapper's test check it against the wrapper's public types, with an explicit +allowlist for items a wrapper deliberately hides. Kotlin's fixed list moves +to the shared one. -## Waiting +Decided 2026-11-15 (maintainer, /quest-complete): ✅ build the list only from +the exporter, so this waits on it (rejected: from UniFFI's metadata now, +rebuilt once the exporter lands). The exporter's switch keeps every +wrapper's API, so a UniFFI version would be written twice. -- Maintainer: when to build the FFI item list: from UniFFI's metadata now, - rebuilt once [the exporter](/quest/m3/ffi-exporter.md) lands, or only from - the exporter, by making that quest Required here. Recommended: only from - the exporter, since its switch keeps every wrapper's API and the UniFFI - version would be written twice. +## Required + +- [Switch moq's bindings to rust2's exporter](/quest/m3/ffi-exporter.md) - the exporter the FFI item list is generated from diff --git a/quest/m4/browser-reset-reports.md b/quest/m4/browser-reset-reports.md index 9fa5ced9783..09959e27c52 100644 --- a/quest/m4/browser-reset-reports.md +++ b/quest/m4/browser-reset-reports.md @@ -39,8 +39,8 @@ both bugs within a few milliseconds of the stream's data. Decided 2026-11-13 (maintainer, `/quest-complete`): ✅ keep this follow-up (rejected: drop it and live with the workaround alone). -## Waiting - -- Maintainer: approve filing at crbug.com and bugzilla.mozilla.org, outside the - moq orgs, which needs your go-ahead; check each tracker's policy on - LLM-written reports first. Recommendation: file both, with the repro. +Decided 2026-11-15 (maintainer, /quest-complete): ✅ approved to file at +crbug.com and bugzilla.mozilla.org, each only after reading that tracker's +policy on LLM-written reports and searching for an existing report. Where a +tracker bans LLM-written reports, hand the maintainer a draft with the repro +instead of filing it. diff --git a/quest/m4/intra-refresh/README.md b/quest/m4/intra-refresh/README.md index ca268987a97..7b9d88b4a93 100644 --- a/quest/m4/intra-refresh/README.md +++ b/quest/m4/intra-refresh/README.md @@ -26,9 +26,14 @@ broadcast feeds that only do intra refresh. Decisions the quests share: -The recovery-boundary planning prerequisite checks the duration contract -below before implementation. Its precision and publication ordering must -make these presentation guarantees true even when sweep lengths change. +Decided 2026-11-15 (maintainer, /quest-complete): ✅ a catalog `warmup` is a +hint nothing enforces (rejected: refusing the stream once a sweep exceeds the +published `warmup`; a bound published before the content it covers; a fixed +upper bound from importer config). Since +[moq-dev/moq#8857](https://github.com/moq-dev/moq/pull/8857) plays a `warmup` +rendition through one decoder, the value only tells a player how early to +join and how long to hide pictures, so a sweep longer than published shows +refresh stripes a little longer: a cosmetic cost, not a correctness one. - One group per refresh cycle. A group starts at the recovery point, the frame that begins a sweep, so a viewer joining at any group boundary is clean after diff --git a/quest/m4/intra-refresh/consumer-warmup.md b/quest/m4/intra-refresh/consumer-warmup.md index 37797e8ba06..15c16b0daf5 100644 --- a/quest/m4/intra-refresh/consumer-warmup.md +++ b/quest/m4/intra-refresh/consumer-warmup.md @@ -31,7 +31,12 @@ The fix is history rather than a withhold rule: - History: before decoding group N, decode the K preceding groups as decode-only input, where K is the fewest groups whose span covers `warmup`. - With one group per sweep that is one group. The history is fed to the + With one group per sweep that is one group. `warmup` is a hint nothing + enforces (see the [epic](/quest/m4/intra-refresh/README.md)), but a stale + value still leaves the history clean: each group opens at a recovery point, + and the [H.264 import](/quest/m4/intra-refresh/h264-import.md) refuses a + sweep whose recovery picture lands past the next one, so one whole + preceding group recovers the decoder. The history is fed to the decoder and never forwarded to the encoder, so group N's first picture is recovered. A group that opens on a true IDR needs no history: check the first slice NAL type (H.264 type 5; H.265 `IDR_W_RADL`, `IDR_N_LP`, the BLA types) @@ -51,7 +56,3 @@ The fix is history rather than a withhold rule: feed start and a feed skip both decode history before forwarding; an IDR group needs no history; a rendition without `warmup` fetches and reads exactly as today. - -## Required - -- [Plan: recovery boundaries](/quest/m4/intra-refresh/recovery-boundary.md) - K derives from `warmup`, which must be a safe upper bound diff --git a/quest/m4/intra-refresh/h264-import.md b/quest/m4/intra-refresh/h264-import.md index 9895ff8acc3..b22059f5e73 100644 --- a/quest/m4/intra-refresh/h264-import.md +++ b/quest/m4/intra-refresh/h264-import.md @@ -11,10 +11,10 @@ were immediately decodable. ## Plan -The recovery-boundary planning prerequisite must settle the precision and -publication ordering before implementing the duration proposal below. Raising -a catalog value after a longer sweep is not yet proven safe for a viewer that -already received those pictures. Decided 2026-10-22 (maintainer): precision +Decided 2026-11-15 (maintainer, /quest-complete): a catalog `warmup` is a +hint nothing enforces (see the [epic](/quest/m4/intra-refresh/README.md)), so +raising it after a longer sweep needs no ordering against the groups it +covers. Decided 2026-10-22 (maintainer): precision follows the hang draft, which rounds up ([#5212](https://github.com/moq-dev/moq/pull/5212)). @@ -39,13 +39,19 @@ follows the hang draft, which rounds up sweep has measured it; a later, longer sweep mutates it upward. A stream whose recovery picture never arrives before the next recovery point is refused, not published with a guess. +- Review the H.264 recovery-point flags and count semantics, and state which + profiles the import supports without changing an existing wire contract by + inference. A timestamp-before-keyframe rule is a conservative + leading-picture policy, not a complete codec dependency model. +- Test timestamp precision, reordered pictures, `frame_num` wrap, variable + sweep lengths, a missed recovery picture, and a join or skip before a longer + sweep is known. Decode-only pictures still build reference state: + withholding presentation is not dropping their input. Use generated + fixtures, run the short cases in CI, and keep full reference-decode + comparisons for continuous playback and each recovery case. - Fixture: a synthetic fixed-framerate access-unit sequence with `recovery_frame_cnt = 2` beside the existing recovery-point tests in `split.rs`, asserting the exact `warmup`, in both the Annex B and the length-prefixed path; and a real clip generated with x264 `intra-refresh=1` added as a second round-trip in `test/ts/run.sh` next to the closed-GOP one, asserting the catalog `warmup` and the group count. - -## Required - -- [Plan: recovery boundaries](/quest/m4/intra-refresh/recovery-boundary.md) - resolve the later-mutated warmup contract before implementing it diff --git a/quest/m4/intra-refresh/h265-import.md b/quest/m4/intra-refresh/h265-import.md index 8a0253e2025..327c5c678e3 100644 --- a/quest/m4/intra-refresh/h265-import.md +++ b/quest/m4/intra-refresh/h265-import.md @@ -31,6 +31,10 @@ that machinery in parallel. recovery point's plus the count and measures `warmup` from its timestamp exactly as the H.264 quest does, sharing the measurement and catalog mutation. +- Review HEVC CRA, RADL, and RASL separately from H.264's recovery points. + The TS export already keeps RADL pictures by NAL type, and + `rs/moq-mux/src/container/ts/recovery_test.rs` grades actual group skips + and recovered output for H.264 open GOP and HEVC RASL and RADL. - Fixtures: synthetic NAL sequences beside the existing suffix-SEI tests, since x265's intra refresh emits no recovery-point SEI, and a redistributable externally encoded clip when available. An NVENC-generated diff --git a/quest/m4/intra-refresh/recovery-boundary.md b/quest/m4/intra-refresh/recovery-boundary.md deleted file mode 100644 index 764fa7dd239..00000000000 --- a/quest/m4/intra-refresh/recovery-boundary.md +++ /dev/null @@ -1,66 +0,0 @@ -# [S] Plan: recovery boundaries remain safe as refresh duration changes - -## Goal - -Settle how consumers know a group's precise recovery boundary before -presenting pictures, including when later refresh sweeps take longer than -earlier ones. Rewrite the import and consumer quests around that contract. - -## Plan - -Decided 2026-11-08 (maintainer, /quest-plan): intra refresh is not part of -this cut, so `warmup` ships with its current duration contract and this plan -settles only what the importer needs. - -Moved to m4 2026-11-12 (maintainer, /quest-audit): no downstream application -waits on intra-refresh import, and no native publisher sets video `warmup`. -The boundary stays a catalog field, never a LOC frame property or other -per-frame container field, and it is a duration, not a frame count. A -per-group exact boundary carried in the media was offered and declined. -Re-interview the maintainer before choosing how a catalog duration stays -safe when a later sweep is longer, since a catalog update and the groups -it covers travel on separate tracks with no ordering between them. - -The 2026-10-11 TS audit found the proposed catalog duration is not yet a -sufficient guarantee. `h264-import` raises the global value after observing -a longer sweep. -`consumer-warmup` sizes the transcoder's decode history from that duration. A -later catalog update cannot -recall pictures a subscriber has already displayed using the old value. - -Decided 2026-10-22 (maintainer): precision is settled; the hang draft rounds a -fractional-millisecond `warmup` up -([#5212](https://github.com/moq-dev/moq/pull/5212)). The draft's `warmup` rule -is amended so a consumer running a parallel max-delay decoder MAY present -min-delay pictures before the group start plus `warmup`, which keeps the -merged dual decoder in `decode::Consumer` -([#5842](https://github.com/moq-dev/moq/pull/5842)), and the draft now says -so. Publication ordering for a later, longer sweep remains to settle. - -Ask the maintainer how the catalog duration stays safe for a longer sweep, -such as a bound published before the affected content or a fixed bound from -config, including the latency each costs. Test timestamp precision, reordered -pictures, frame_num wrap, variable sweep lengths, a missed recovery picture, -and a join/skip before a longer sweep is known. Decode-only pictures still -build reference state; withholding presentation is not dropping their input. - -Review the H.264 recovery-point flags and count semantics and HEVC -CRA/RADL/RASL separately. A timestamp-before-keyframe rule is a conservative -leading-picture policy, not a complete codec dependency model; the TS export -already keeps HEVC RADL pictures, by NAL type, and -`rs/moq-mux/src/container/ts/recovery_test.rs` grades actual group skips and -recovered output for H.264 open GOP and HEVC RASL and RADL. State which -profiles it supports without changing an existing wire contract by inference. - -Produce implementation and automated software-decode tests in the existing -quests. Use generated fixtures, run short cases in CI, and retain full -reference-decode comparisons for continuous playback and each recovery case. - -## Waiting - -- Maintainer: how a catalog `warmup` stays safe when a later sweep is longer than the published one. Options: a bound published before the content it covers (costs one sweep of latency up front); a fixed upper bound from importer config; refusing the stream once a sweep exceeds the published `warmup`. Recommended: refuse, matching h264-import's rule that a stream is refused rather than published with a guess, and let a configured bound raise the limit where an operator knows its encoder. - -## Related - -- [H.264 import](/quest/m4/intra-refresh/h264-import.md) - publishes recovery information -- [Consumer warmup](/quest/m4/intra-refresh/consumer-warmup.md) - uses it to size the transcoder's decode history diff --git a/quest/m3/quic-ack-hook.md b/quest/m4/quic-ack-hook.md similarity index 66% rename from quest/m3/quic-ack-hook.md rename to quest/m4/quic-ack-hook.md index d41acfa7b6c..01d4977d31c 100644 --- a/quest/m3/quic-ack-hook.md +++ b/quest/m4/quic-ack-hook.md @@ -1,10 +1,10 @@ -# [S] The QUIC drivers' acknowledgment wait is public and tested at its traps +# [S] The QUIC drivers' acknowledgment wait is tested at its traps ## Goal -moq-tokio's and moq-uring's QUIC drivers let a caller await acknowledgment -of a minted `send::Offset` on `send::Stream` and `send::Finished`, publicly, -and each driver's tests cover the traps that wait has. +moq-tokio's and moq-uring's QUIC drivers each have tests covering the traps +of their crate-private acknowledgment wait, the one their `quic::web` +adapters answer `poll_acked` with. The wait stays crate-private. ## Plan @@ -21,17 +21,13 @@ trait's default `Acked::Unsupported` (decided 2026-10-22 and 2026-10-26, maintainer): writing into TCP is not delivery, and the browser's `WebTransportSendStream.getStats()` is unimplemented in shipping Chrome. -What is left is making them public, in the drivers' caller-owned waiter shape -with an async wrapper: - -```rust -impl send::Stream { - pub fn poll_acked(&self, waiter: &kio::Waiter, offset: send::Offset) -> Poll>; - pub async fn acked(&self, offset: send::Offset) -> Result<(), send::Error>; -} -``` - -The same pair, and `offset()`, go on `send::Finished`. +Decided 2026-11-15 (maintainer, /quest-complete): ✅ keep only the trap tests, +on the crate-private methods, and drop the public driver API this quest +planned (`poll_acked`, `acked`, and `offset()` on `send::Stream` and +`send::Finished`) (rejected: make the wait public; delete the quest). Since +[moq-dev/moq#8845](https://github.com/moq-dev/moq/pull/8845) the adapters +answer `poll_acked`, so nothing needs the drivers' wait public. Moved from m3 +to m4 the same day (maintainer): it is now tests only (rejected: keep it in m3). Traps to test through each driver: @@ -55,14 +51,7 @@ The wait carries no instant: a caller records one when woken, which includes the peer's delayed-ACK timer (up to `max_ack_delay`) as well as the return trip (decided 2026-10-17 on #5817). -## Waiting - -- Maintainer: whether the drivers' wait goes public at all. Nothing in the - tree needs it public since moq-net reaches it through the adapters, and the - rules keep an API private until a consumer needs it. Options: (1) keep the - quest, if an out-of-tree driver user is expected to await - acknowledgments; (2) recommended: drop the public half and keep only the - trap tests, on the crate-private methods, resizing this quest to XS. +Public API: none. Wire: none. ## Related diff --git a/quest/m4/quic-deadline.md b/quest/m4/quic-deadline.md index df3a3eb7134..44876e2c5b3 100644 --- a/quest/m4/quic-deadline.md +++ b/quest/m4/quic-deadline.md @@ -36,7 +36,7 @@ Implement in `moq-quic`. the congestion window; the probe is a scheduling choice, not extra credit. - moq-net cannot call `moq-quic` directly, so the deadline crosses its own `transport::poll::SendStream` trait, the way - [the ACK hook](/quest/m3/quic-ack-hook.md) does: a method whose default + [the ACK hook](/quest/m4/quic-ack-hook.md) does: a method whose default reports unsupported, implemented by the moq-tokio and moq-uring adapters over `moq-quic`. A backend without it sends as today. - moq-net: the per-group `GroupServe` machine in the lite and IETF @@ -101,4 +101,4 @@ measured. ## Related -- [ACK hook](/quest/m3/quic-ack-hook.md) - the same pattern for a moq-net method backed by `moq-quic` +- [ACK hook](/quest/m4/quic-ack-hook.md) - the same pattern for a moq-net method backed by `moq-quic` diff --git a/quest/m4/quic-interop-listing.md b/quest/m4/quic-interop-listing.md index 7a9aa9df8ef..6eda91ab3c4 100644 --- a/quest/m4/quic-interop-listing.md +++ b/quest/m4/quic-interop-listing.md @@ -13,10 +13,11 @@ listing; listing from day one). The registration is a PR outside the maintainer's orgs, so it needs their go-ahead, and the repository's AI policy is read before drafting it. +Decided 2026-11-15 (maintainer, /quest-complete): ✅ approved to register on +the public matrix once every supported case passes against at least two +peers in our nightly run. Where the AI policy above bans LLM-written PRs, +hand the maintainer a draft instead. + - Publish the endpoint image `test/quic-interop` builds, from `main`, and add the entry to `implementations_quic.json`. Public API: none. Wire: none. - -## Waiting - -- Maintainer: approve registering on the public matrix once the nightly matrix is mostly green. An agent cannot decide when our failures become public or post to quic-interop unasked. Recommendation: approve when every supported case passes against at least two peers. diff --git a/quest/m4/resume-redesign.md b/quest/m4/resume-redesign.md deleted file mode 100644 index 794487ac1a6..00000000000 --- a/quest/m4/resume-redesign.md +++ /dev/null @@ -1,110 +0,0 @@ -# [XL] Resume as producer replacement on one track - -## Goal - -A route change under a front replaces the producer writing into the -front's one logical track state, instead of every reader merging route -copies through its own cursors. Readers become plain cursors, the -`Inner::Resume` path and its per-reader copy bookkeeping are deleted in Rust -and `js/net`, and a benchmark swept over readers and route flaps shows the -cost no longer grows with readers times copies. - -## Plan - -From the 2026-10-24 pre-cut audit at 151a529670. Internal only: no public API -or wire change, so it can ship after the cut. The audit recommended after -the cut, not before: a rewrite now would discard the large resume test -corpus right before release. Decided 2026-10-25 (maintainer): the -model test (`rs/moq-net/src/model/resume_model.rs`) came first and gated the -cut; this rewrite moves to m1 only if that test shows design-level bugs. -It found single missed checks -([moq-dev/moq#6947](https://github.com/moq-dev/moq/pull/6947)), so this -stays after the cut, with the model as its oracle. - -**Why.** Main replaced release's splice design with per-reader route copies -in 0382d309f4 (2026-10-04). Since then about 30 of 58 commits to -`rs/moq-net/src/model/resume.rs` were fixes, rising to about 10 a week, and -production code grew from 1,004 to 1,989 lines (`track.rs` from 4,203 to -7,626, partly features). The fixes cluster on state one logical reader keeps -in several owners, which must be kept in agreement by hand: - -- where the reader is (`Reader.groups/window`, `Copy.floor/from/resolved`, - cursor bounds, `Delivered.floor`): fd59eb18a5, 9f1fc92fe3, 35fa44f6bf, 81d8a6313c, 137de79100 -- what was handed out (logical and per-copy `delivered`, `Reader.newest/out`, - the two-level peek): 323fda277f, bfbaa0b46d, 342445f251, 3602a8af9a -- whether a replaced copy is still needed (`Copy::retains` asking another - layer's cursor): 9b1b81a496, 5e6c452305, 9bba359f6b, 2fbaa59428 -- route failure versus content verdict: 39eacc807a, 109235732c, eb6ac57cb7, 5c31208409 -- end agreement and budget/expiry: 25532eacc8, a602360c99, 097dc1ee62, f52848d634, b59fcefde3, e33fe24cc3 - -The churn continues: `resume.rs` took 28 commits from 2026-11-01 to -2026-11-12, 9 of them fixes (for example e419224dc, f3edb0b8d). Since this -was written, `rs/moq-net/verify/resume/ResumeReader.tla` and -`rs/moq-net/verify/failover/FailoverChain.tla` model the per-reader path; a -redesign replaces or re-maps them (the `verify-resume` and `verify-failover` -skills name the code each covers). - -Every `track::Consumer`/`Subscriber` method has two implementations (26 -`Inner::Resume` arms and 11 `serving()` delegations then; 45 arms and 26 -`serving()` calls in `track.rs` at 4bf501b535), so each feature (ranges, -anchors, fills, compression, largest, drops, start, live) is written twice -and drifts. `js/net/src/resume.ts` (1,786 lines then, 2,296 at 4bf501b535) -mirrors it, and the resume floor fix -[moq-dev/moq#6759](https://github.com/moq-dev/moq/pull/6759), the resumed JS drops fix -[moq-dev/moq#6758](https://github.com/moq-dev/moq/pull/6758), and the compress check -[moq-dev/moq#6460](https://github.com/moq-dev/moq/pull/6460) are each a place the two -paths diverged. - -**Sketch.** - -1. Route copies become successive writers into the logical `TrackState`, - the only cache. Every writer is of the instance the subscription started - on: a newer [publisher epoch](/doc/concept/moq-lite.md#publisher-instances) - never feeds it (the route rule in `rs/moq-net/AGENTS.md`), so its readers - end with an error once their instance's routes are gone. Same-epoch routes - carry identical content, so a later writer's duplicate sequence is - dropped or kept as the slot's backup. The old writer keeps feeding below - the new writer's first group and is released once the logical cache - covers that span or the aggregate budget passes, judged once per track by - `drifted`, not per reader. -2. The upstream request is already the union of readers' resume points - (`Producer::subscription`), so per-reader copy cursors are local overhead. -3. Readers become plain cursors: delete `Inner::Resume`, `Reader`, `Copy`, - `Retain`, `Candidate`, and per-reader `learn_end`. -4. `Recover` attaches to the slot, not a handed-out group: continuing a - group whose copy failed is a - [flight](https://github.com/moq-dev/moq/pull/7944) from the reader's frame on - the new route, which hands over like any other flight. - -Hard constraint from the splice era: two caches sharing one group handle -caused expiry bugs (f9b7982e01, 6c2b733bbb) and boundary churn (06bcadd06c, -b9398f7459, 68c73349a3). Session copies under a front must stop caching for -themselves and only feed the logical track. - -Land it in Rust first, then mirror `js/net` unless -[the Rust model in the browser](/quest/m5/rs2ts-js.md) has replaced it. Extend -`benches/origin.rs` with the reader times flap sweep before starting, to -measure the change. - -Nightly mutates `resume.rs` (#8712): the rewrite kills or excludes, in -`.cargo/mutants.toml`, every survivor of the code it adds, and drops the -exclusions of the code it deletes. - -Public API: none. Wire: none. - -Decided 2026-11-05 (maintainer, follow-up from -[moq-dev/moq#7736](https://github.com/moq-dev/moq/pull/7736)): ✅ fold this -case into the redesign instead of patching code it deletes. In -`js/net/src/resume.ts`, `#update` skips a copy whose upstream subscription -closed, so a closed copy that still holds groups keeps picking under the -reader's old `delay.max` after an update. It was found by review and not -reproduced. The redesign's tests must cover a preference update after the -serving copy's upstream closed, followed by a read. - -## Waiting - -- Maintainer: whether this goes before [resume trace validation](/quest/m4/resume-trace-validation.md), and whether it moves to m3 once the cut ships. Mutation testing already covers the current `resume.rs` ([moq-dev/moq#8712](https://github.com/moq-dev/moq/pull/8712)). Options: (a) redesign first, giving trace validation `Required` on this quest, and move this to m3 after the cut as the root fix for the resume path's recurring bugs in shipped code; (b) validate the current path first, keeping this in m4; (c) redesign first but keep it in m4. Recommended: (a), since validation would spend its effort on the per-reader path this deletes. - -## Required - -- [Release cut](/quest/m3/release/cut.md) - wait for the release in flight to ship diff --git a/quest/m4/resume-trace-validation.md b/quest/m4/resume-trace-validation.md index fb854733727..06c1f8731cf 100644 --- a/quest/m4/resume-trace-validation.md +++ b/quest/m4/resume-trace-validation.md @@ -25,16 +25,11 @@ can keep passing on code that has moved on. Milestone m4. - The sweep covers more than the model does (ranges, cache expiry, a second instance). Restrict the validated cases to what the model covers or grow the model, whichever costs less, and say which. -- Before starting, check [Resume as producer replacement](/quest/m4/resume-redesign.md): - it deletes the per-reader path this model describes. If it has landed or is - in flight, validate against the redesign's model instead, or delete this - quest with the old one. +- The per-reader path this model describes stays: the producer-replacement + redesign that would have deleted it was dropped 2026-11-15 (maintainer, + /quest-complete). - Nightly on the sweep's fresh seed (`test resume-model`); the merge gate only if it stays fast on the fixed cases. Public API: none. Wire: none. - -## Related - -- [Resume as producer replacement](/quest/m4/resume-redesign.md) - rewrites the code this validates diff --git a/quest/m4/upstream-reports.md b/quest/m4/upstream-reports.md index 1a0f702734c..942a0d5236d 100644 --- a/quest/m4/upstream-reports.md +++ b/quest/m4/upstream-reports.md @@ -1,4 +1,4 @@ -# [S] Report upstream bugs found during the pre-cut run +# [S] Report approved upstream bugs ## Goal @@ -7,18 +7,14 @@ our workaround links the report. ## Plan -Posting outside `moq-dev` and `kixelated` needs maintainer approval for each -report. - -- **GStreamer `faad`:** it drops the first AAC frame and shifts timestamps by - one frame, so the interop AAC phase check skips GStreamer - ([moq-dev/moq#6742](https://github.com/moq-dev/moq/pull/6742)). Remove the - skip once it's fixed. -- **retina 0.4.21:** when the server sends an RTSP request during playback, - retina logs a warning, returns `Pending`, and goes idle until something else - wakes it ([moq-dev/moq#6736](https://github.com/moq-dev/moq/pull/6736)). -- **mr-boxington:** the eviction found while landing - [moq-dev/moq#6709](https://github.com/moq-dev/moq/pull/6709). +Decided 2026-11-15 (maintainer, /quest-complete): ✅ (b) drop the three +reports that had no recorded approval, the GStreamer `faad` first-frame drop +([moq-dev/moq#6742](https://github.com/moq-dev/moq/pull/6742)), retina +0.4.21's idle `Pending` +([moq-dev/moq#6736](https://github.com/moq-dev/moq/pull/6736)), and the +mr-boxington eviction ([moq-dev/moq#6709](https://github.com/moq-dev/moq/pull/6709)), +and keep only the approved reports below (rejected: (a) approve some or all). +Each workaround already works and none blocks other work. Added 2026-10-31 (user, /quest-plan of the follow-ups from the 2026-10-31 /quest-complete run). Upstream reports: ✅ Chromium VideoFrame leak @@ -29,7 +25,9 @@ B-frames realtime ([moq-dev/moq#7402](https://github.com/moq-dev/moq/pull/7402)) file; str0m/dimpl DTLS budget ([moq-dev/moq#7379](https://github.com/moq-dev/moq/pull/7379)) NOT approved. -Approved to file (search for an existing report first): +Approved to file, each after searching for an existing report and reading +the tracker's policy on LLM-written reports; where it bans them, hand the +maintainer a draft instead: - **Chromium:** tearing down a `MediaStreamTrackProcessor` with a frame still in flight inside it drops that `VideoFrame` unclosed, so Chromium @@ -57,7 +55,3 @@ Not to be filed: [DTLS waits for ICE](/quest/m3/webrtc/dtls-after-ice.md) fixes it there. Public API: none. Wire: none. - -## Waiting - -- Maintainer: which of the reports with no recorded approval may be posted: the GStreamer `faad` frame drop, retina 0.4.21's idle `Pending`, and the mr-boxington eviction (which [Mbx R2 verify](/quest/m0/ci/mbx-r2-verify.md)'s upstream threads may already cover). Options: (a) approve some or all, each after reading that project's AI policy and searching for an existing report; (b) drop all three and keep the approved Chromium and Firefox reports. Recommended: (b), since each workaround already works and none blocks other work. diff --git a/quest/m4/windows-gpu-tests.md b/quest/m4/windows-gpu-tests.md new file mode 100644 index 00000000000..2711ca9492a --- /dev/null +++ b/quest/m4/windows-gpu-tests.md @@ -0,0 +1,48 @@ +# [M] The Windows GPU tests run on a real GPU + +## Goal + +The ignored moq-video tests that need a Windows GPU, and that no CI runner +hosts, pass on the maintainer's `windows` host (Windows 11, NVIDIA RTX 2070), +and what a run breaks is fixed. + +## Plan + +Split from [Video hardware validation](/quest/m5/video-hardware.md) 2026-11-15 +(maintainer, /quest-complete): these items need only that host's GPU, not the +cameras and other machines the rest of that quest waits on. Running ignored +tests changes no API, but a downstream studio's Windows run lists the parent +quest as related, so they sit in m4 rather than the lab. + +- Direct3D11 video processing: `d3d11_resize_defaults_to_the_gpu`, + `d3d11_resize_matches_cpu`, and + `mediafoundation_resized_texture_reencodes_in_place`. `VideoProcessorBlt` + can hang on affected drivers, so run them one at a time. +- WGC conversion and encode: `wgc_conversion_crops_odd_edges_and_preserves_color`, + `wgc_conversion_workload`, and + `wgc_nv12_encodes_with_matching_color_on_gpu_and_software`. The `wgc_*` + capture tests belong to [Windows.Graphics.Capture](/quest/m3/capture-wgc.md). +- Media Foundation encode: `mediafoundation_cpu_rgba`. + `mediafoundation_camera_texture` needs a camera the host lacks, so it stays + in the parent quest. +- A GPU export on a hardware encoder: + `render::export::dx12::tests::exports_an_nv12_texture` (the Platform runner + has only WARP, which has no Direct3D11 video device) and + `render::export::dx12::tests::media_foundation_encodes_the_export_in_place`. + The [studio](https://github.com/moq-dev/studio), the export's main consumer, + confirms in its own repository that its Output panel reports no readback on + this machine. + +Fix only a defect a run hits, and record the driver version each test ran on. +Running them here gives them no CI caller, so `sh/gh/test-inventory.allow` +keeps naming the parent quest as their owner. + +Public API: none. Wire: none. + +## Hosts + +- windows - a Windows GPU with Direct3D11 video processing and a Media Foundation hardware H.264 encoder + +## Related + +- [Video hardware validation](/quest/m5/video-hardware.md) - the hardware paths that still wait on cameras and other machines diff --git a/quest/m5/video-hardware-access.md b/quest/m5/video-hardware-access.md index 10c0f057c6c..92a28fb3865 100644 --- a/quest/m5/video-hardware-access.md +++ b/quest/m5/video-hardware-access.md @@ -33,8 +33,9 @@ Record the device, OS image, and driver when one is on hand. Checked 2026-11-12 (/quest-audit): part of the Windows condition holds. The maintainer's `windows` host (Windows 11, NVIDIA RTX 2070) has hosted agents natively since 2026-11-10, so the Windows GPU probes and the Media Foundation -encode of a GPU export can run there. It has no camera, so the MJPEG and YUY2 -camera items still wait. Whether the maintainer's macOS laptop is Apple Silicon +encode of a GPU export were split out to run there, as +[Windows GPU tests](/quest/m4/windows-gpu-tests.md) (2026-11-15). It has no +camera, so the MJPEG and YUY2 camera items still wait. Whether the maintainer's macOS laptop is Apple Silicon is not recorded; ask. The validation can start with whatever subset is available; split the rest diff --git a/quest/m5/video-hardware.md b/quest/m5/video-hardware.md index b508fd7df53..0efc13d1887 100644 --- a/quest/m5/video-hardware.md +++ b/quest/m5/video-hardware.md @@ -18,18 +18,21 @@ Meteor Lake nor the Arrow Lake test host exposes `EncSliceLP`, so no machine on hand reaches that fallback. A render node other than the first is checked by [VA-API external images](/quest/m5/vaapi-vulkan-import.md). +Decided 2026-11-15 (maintainer, /quest-complete): ✅ split the items the +`windows` host (RTX 2070, no camera) can run today, the Windows GPU probes and +the two dx12 export tests, into [Windows GPU tests](/quest/m4/windows-gpu-tests.md) +(rejected: splitting them into the lab; keeping everything here until the +whole hardware set is on hand). + - **Windows Media Foundation capture**: the camera opens while capture is enabled and closes when it is disabled, so the LED is off whenever capture is; with the `on_demand` opt-in it also opens and closes as the first viewer - arrives and the last leaves. NV12 delivery from MJPEG and YUY2 cameras. -- **A GPU export on a hardware encoder**: run the ignored - `render::export::dx12::tests::exports_an_nv12_texture` (the Platform runner - has only WARP, which has no Direct3D11 video device) and - `render::export::dx12::tests::media_foundation_encodes_the_export_in_place` - on a Windows GPU with a Media Foundation H.264 encoder. The - [studio](https://github.com/moq-dev/studio), the export's main consumer, - confirms its Output panel reports no readback on that machine and a Mac in - its own repository. + arrives and the last leaves. NV12 delivery from MJPEG and YUY2 cameras, + and the ignored `mediafoundation_camera_texture` encode test, which needs + a camera and a GPU. +- **A GPU export on a Mac**: the [studio](https://github.com/moq-dev/studio), + the export's main consumer, confirms in its own repository that its Output + panel reports no readback on an Apple Silicon Mac. - **A live camera run per platform**: capture needs device permission that a headless or agent process cannot grant itself. - **OBS follows its grant on a shaped uplink**: stream from a real OBS build @@ -70,22 +73,6 @@ A separate-plane producer is not planned; if that is why a Pi produces nothing, write it down and stop. `doc/lib/rs/moq-video.md` says both paths are reachable; correct that sentence if one cannot capture. -### Windows GPU probes - -Ignored moq-video tests that only a GPU on that platform can run, and that no -runner hosts (found by the m0 coverage audit): - -- Direct3D11 video processing: `d3d11_resize_defaults_to_the_gpu`, - `d3d11_resize_matches_cpu`, and - `mediafoundation_resized_texture_reencodes_in_place`. `VideoProcessorBlt` - can hang on affected drivers, so run them one at a time. -- WGC conversion and encode: `wgc_conversion_crops_odd_edges_and_preserves_color`, - `wgc_conversion_workload`, and - `wgc_nv12_encodes_with_matching_color_on_gpu_and_software`. The `wgc_*` - capture tests belong to [Windows.Graphics.Capture](/quest/m3/capture-wgc.md). -- Media Foundation encode: `mediafoundation_cpu_rgba` and - `mediafoundation_camera_texture`. - ### Fixes Fix only a defect a run hits. Precedent for what this catches: NVENC @@ -93,15 +80,12 @@ validation on an RTX 3070 Ti found that NVENC rejects stream-ordered pool memory, so buffers registered with it must come from plain `cuMemAlloc`. That is not a bug any amount of review finds. -## Waiting - -- Maintainer: whether to split the items the `windows` host (RTX 2070, no camera; see [hardware access](/quest/m5/video-hardware-access.md)) can run today, the Windows GPU probes and the two dx12 export tests, into their own quest with `Hosts: windows`. Options: split them into m4 (or into lab, since running ignored tests changes no API), or keep everything here until the whole hardware set is on hand. Recommendation: split into m4 with `Hosts: windows`, since a downstream studio's Windows run lists this quest as related and those tests need nothing else. - ## Required - [Video validation hardware is on hand](/quest/m5/video-hardware-access.md) - the machines to run it on ## Related +- [Windows GPU tests](/quest/m4/windows-gpu-tests.md) - the Windows GPU probes and dx12 export tests split out to run on the `windows` host - [#2819](/quest/m4/2819-moq-video-carry-pipewire-dma-bufs-safely-into-the-vulkan.md) - the DMA-BUF import validation this capture feeds - [Embedded video path](/quest/m5/video-embedded.md) - presenting on a Pi, which is a different gap diff --git a/quest/m9/stream-transports/relay.md b/quest/m9/stream-transports/relay.md index 846de7f1986..fe5373ec7f3 100644 --- a/quest/m9/stream-transports/relay.md +++ b/quest/m9/stream-transports/relay.md @@ -28,10 +28,18 @@ transport is driven, which is what makes that handoff free here too. moq-tokio spawns each qmux `Driver` on its own runtime (`spawn_qmux` in `rs/moq-tokio/src/lib.rs`), so its `tcp::Listener::accept`, `websocket::Listener::accept`, and `websocket::accept` return only the session. -A worker must run the driver on its own local task set instead, so this quest -reshapes `moq_tokio::websocket::accept` (or adds a sibling) to hand the driver -back; the qmux constructors already return it (decided 2026-11-01, maintainer, -in [moq-dev/moq#7494](https://github.com/moq-dev/moq/pull/7494)). +A worker must run the driver on its own local task set instead, so +`moq_tokio::websocket::accept` must hand the driver back; the qmux +constructors already return it (decided 2026-11-01, maintainer, in +[moq-dev/moq#7494](https://github.com/moq-dev/moq/pull/7494)). + +Decided 2026-11-15 (maintainer, /quest-complete): ✅ reshape +`moq_tokio::websocket::accept` itself to return the session and its driver +(rejected: an additive sibling beside it). That breaks moq-tokio's published +API, which the lab never touches, so ✅ the reshape is its own m2 quest, +[WebSocket accept hands back the driver](/quest/m2/websocket-accept-driver.md), +listed in Required (rejected: landing it here with the rust2 migration). +This quest keeps only the relay's internal half. Keep the ops and web listeners' behavior identical: the same routes, the same CORS scoping, the same landing-page fallback, and the same @@ -39,11 +47,8 @@ CORS scoping, the same landing-page fallback, and the same to prove a WebSocket session and an HTTP route both work when served from a worker. -## Waiting - -- Maintainer: how moq-tokio hands a WebSocket session's driver back, since this lab quest must not change a published API without a decision. Options: add an additive sibling beside `moq_tokio::websocket::accept` that returns the session and its driver, or reshape `websocket::accept` itself, which breaks moq-tokio and would have to land before the cut. Recommendation: the additive sibling, so nothing here waits on the cut. - ## Required +- [WebSocket accept hands back the driver](/quest/m2/websocket-accept-driver.md) - the moq-tokio API that returns the qmux driver for a worker to run - [Ring TCP](/quest/m9/stream-transports/ring.md) - the module and adapters this serves from