Repository navigation
fix: resolve the fork-only issues on PR #3 - #57
Merged
Merged
Conversation
Cover every Bun workspace with the scope regression check. Fixes #48. Co-Authored-By: Codex <noreply@openai.com>
Exercise real replay output paths and probe sampling, and test large recording floors. Each regression fails when its corresponding safeguard is reverted. Covers the harness portion of #50. Co-Authored-By: Codex <noreply@openai.com>
Reject zero report intervals at argument parsing. Count datagrams by active profile without adding a lock or an extra atomic update to forwarding. Cover stepped impairments and benchmark clients against step counts in nightly CI. Fixes #47. Co-Authored-By: Codex <noreply@openai.com>
Refs #31. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Co-Authored-By: Codex <noreply@openai.com>
Preserve decoded audio across bounded timeline holes. Commit short fragments before the splice instead of discarding them, and drain finite tracks without waiting for an impossible refill. Refs #32 and #33. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Co-Authored-By: Codex <noreply@openai.com>
Refs #39. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Co-Authored-By: Codex <noreply@openai.com>
Refs #45. Verified raw QUIC and WebTransport admission against Linux io_uring workers. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Co-Authored-By: Codex <noreply@openai.com>
Fixes #31 for the JavaScript playout engine. Co-Authored-By: Codex <noreply@openai.com>
Fixes #43. Co-Authored-By: Codex <noreply@openai.com>
Fixes #40. Co-Authored-By: Codex <noreply@openai.com>
Bound worker startup including lazy imports, terminate late workers, and preserve the refusal for later players. Fixes #42. Co-Authored-By: Codex <noreply@openai.com>
Fixes #39E. Co-Authored-By: Codex <noreply@openai.com>
Keep the first playback report and errors on the page while the worker receives ongoing ring state. Fixes #39D. Co-Authored-By: Codex <noreply@openai.com>
Fixes #39A. Co-Authored-By: Codex <noreply@openai.com>
Distinguish codec recovery from the stall watchdog and remove load-sensitive deadlines. Fixes the watch portion of #53. Co-Authored-By: Codex <noreply@openai.com>
Keep preview timestamps out of promotion and preserve healthy sparse subscriptions. Reset learned cadence when the source changes. Fixes #36 and #38A. Co-Authored-By: Codex <noreply@openai.com>
Fixes #38B. Co-Authored-By: Codex <noreply@openai.com>
Fail an interrupted UDP benchmark batch instead of waiting until the nightly job timeout. Co-Authored-By: Codex <noreply@openai.com>
Start guarded writes only while the group is live. Release partial subscriptions when setup fails, and keep rejected signal listeners out of the subscriber set. Preserve lite priority updates through FIN acknowledgment without expiry listeners. Map local expiry and remote OLD resets to Expired, matching Rust and allowing JSON window consumers to resume at checkpoints. DELIVERY_TIMEOUT remains a distinct stream error. Co-Authored-By: Codex <noreply@openai.com>
Apply positive age budgets to completed groups when the reader stops pulling. Preserve zero-budget burst behavior. Report covered marker groups as discontinuities and avoid mistaking sparse arrivals for event-loop stalls under timer throttling. Pin arrival measurement before parsing and preserve a long GOP after a shorter completed group. Co-Authored-By: Codex <noreply@openai.com>
Keep new frames readable when a returning frame callback predates the last worker tick by no more than one frame period. Preserve rejection of larger clock reversals and make the suspended-callback grace test use a controlled clock. Co-Authored-By: Codex <noreply@openai.com>
Co-Authored-By: Codex <noreply@openai.com>
Extend #40 coverage to detach and reattach with a different codec configuration. Co-Authored-By: Codex <noreply@openai.com>
…s it (moq-dev#4104) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…dev#4103) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Grok 4.7 <noreply@x.ai>
An IETF live join delivers nothing below the group its SUBSCRIBE_OK names as Largest, so declare that as the copy's start. A parked track's warm cache waiting on the copy then sees the upstream moved past it and live readers skip it, as they already do on lite-06+. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Use the shared shutdown signal in the player and send the QUIC close after the window exits. Exercise both process signals with real signals in isolated nextest processes. Refs #51. Co-Authored-By: Codex <noreply@openai.com>
fperex
marked this pull request as ready for review
October 1, 2026 16:16
This was referenced Oct 1, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
PR #3 (
1217ea74b) had 24 openfork-onlyissues (#31 to #53, #56). Its Interop job was red on media reattach: the relay served a returning viewer media from before it left (#51). #55, labelled upstream-code, is the player painting that older media over newer pictures, through fork-only reattach code. On Linux its Audio quality lane underran: the original failed 4 of 36 enforced checks. When this PR was handed over (f21282589), its Interop check was red, its last Nightly (84e98b3a8) failed the Linux Audio quality lane, and the relay cost more CPU than the original at 4 publishers x 10 viewers.Approach
The issue fixes include targeted regression tests. #34 closes by decision, and #52 with PR #3's rewritten description. No budget, timeout or duration was relaxed.
Audio playout, JS and Rust: js/watch, rs/moq-audio: the playout noise estimate locks out for about 33 s after a window of digital silence #31, rs/moq-audio: native playout stalls on a run shorter than one block, and a skipped group in a burst wipes the audio already held #32, rs/moq-audio: native playout conceals through a declared pause instead of parking on the endpoint #33, js/watch audio: the SharedArrayBuffer ring's clock is stamped at poll time, not mapped to the output device, so video leads the sound #37, js/watch, rs/moq-mux: low-severity efficiency and latent defects (clock wakes, partial output timestamp, spurious waker, worker-mode state traffic, ring built before timing) #39, js/watch audio: after a worker fallback the decoded sample rate is forgotten and the AudioContext is rebuilt twice #40, js/watch audio worker: after the AudioContext resumes, the worker's output clock sits at the page's time origin for up to 1 s #41, js/watch audio worker: a worker that never says ready is never refused or terminated, so every later player waits out its own 5 s deadline #42, js/watch audio: a replaced render worklet node keeps its processor (and the playout engine) running until the AudioContext closes #43, js/watch: after a <moq-watch> reattach, an output timestamp with no device time puts sync.now() up to 80 s ahead #56.
Video and the container consumer: js/hang consumer: a finished head at the cursor shields every finished group behind it from the age budget while the reader is paused #35, js/watch video: a sparse rendition (static screen share) is torn down and resubscribed at every silence longer than 5.5 s #36, js/watch video and js/hang consumer: four smaller defects (early promotion, renderer pairs and lag, rationed-tick stall flags, lost marker event) #38, and the Firefox capture polyfill (js/publish: on the Firefox capture polyfill, a frame callback stamped just before the last worker tick tears the capture down #49). A reopened or reattached track never paints a picture older than the one shown, and keeps the held picture until it paints its own (js/watch: a video track reopened after a gap paints pictures older than the one on screen #55).
Returning viewers get no stale cache (js/hang consumer: a kept group that never finishes holds delivery for about one age budget, and video then waits for a keyframe #34, PR #3 lacks upstream fixes that explain its red Interop job and three open defects #51): imports upstream fix(net): hold a parked track's warm cache until the upstream confirms it moq-dev/moq#4104, fix(net): judge spliced staleness against the logical live edge moq-dev/moq#4103 and fix(net): skip a stale warm cache on an IETF rejoin moq-dev/moq#4150, and a publisher now declares a break at every pause marker, so lite-06+ and IETF relays skip the cache from before the pause.
Shared finished-cache confirmation:
a3f1a74f8requests the newest finished group from frame 0 on warm confirmation, so another front sharing the upstream track can replay its head. It retransfers that full finished payload, including completed media, while open continuation and the logical splice boundary stay unchanged. A paused-time real-Origin test reproducesLaggedbefore the fix and reads literal snapshot bytes after; all 1,277 moq-net tests passed. No public API or wire-format change.Transport and relay: js/net: the expiry guard held through a group's FIN never resets anything, and its per-group subscriptions trip the dev subscriber cap #44, rs/moq-relay: during a drain the io_uring workers admit a new session and GOAWAY it, where the tokio listeners refuse it with 503 #45, js/net: an expired group is reset with DELIVERY_TIMEOUT (0x2) where Rust sends Old (0x34), and @moq/json treats 0x2 as fatal #46.
moqcloses its QUIC connections on SIGINT and SIGTERM (fix(cli): close the relay connection on SIGINT and SIGTERM moq-dev/moq#4287, plusmoq play).Tests and tooling: rs/moq-shaper:
--report-interval 0spanics, andverifyjudges impairments against the starting profile, never the steps #47, js/justfile: the JS gates skip test/audio-quality/, so a harness-only change runs no type check, lint or test #48, ci: two PR #3 fixes have no test that fails without them (d0296a562 arrival clock, 8d593b17e audio-quality harness) #50, js/watch, js/publish: three PR #3 tests race the real clock and fail under load #53.Linux audio CI:
moq importread stdin 64 bytes at a time, so audio waited behind 200 to 400 KB keyframes. It now reads 64 KiB (ca4817459). The harness pins-readrate_catchup 1, and a silence-share breach passes only when every quiet window is proven quiet in the source.Receiver stall cadence:
ba8cc67cakeeps source-burst cadence when receiver timers are rationed. Stable seven-frame, 162 ms bursts no longer produce false local-stall flags; 371 ms source gaps can still be flagged. After learning 162 ms spacing, a 150 ms freeze with no task pending can be excused; an outstanding task probe directly measures receiver wait. The fix changes no public API or wire format.Fixed AAC source control:
bc2680217gives fixed-250 AAC its own one-frame-per-PES publisher, so the source can sustain that delay. Adaptive AAC retains the seven- and sixteen-frame packed source. The metric ceilings, timeouts and run durations stay unchanged.Owned task retention proof:
df3d2b444weakly counts 10,000 effect-owned native catch-result promises. Explicit WeakRef job-root release and synchronous sweeping preserve the fewer-than-100 bound; the effect closes infinally. All 81 Signals tests, types and lint pass on Bun 1.3.13; a mutant omitting task deletion retains all 10,000 and fails. Production Signals is unchanged. The earlier ARM CI global count of 8,194 extra Promises remains retained and unattributed.Auth expiry test
ab075bbb0. Media roundtrips between 10 s virtual steps keep QUIC active through the unchanged 120 s lease/30 s slack; cadence 100 ms, default timeouts unchanged. Auth default/jemalloc: 12 pass each; default relay: 232 pass/two skip; broader relay: 339 pass/two skip. LiteralExpiredrequired. Prior2a78f7425red (282 pass/1 fail/56 not run, transport idle versusExpired) and DF publisher timeout stay retained. Test-only, no public API/wire change.Dependency audit lock:
75584b85fchanges the yankedyoke-derivelock entry from 0.8.3 to 0.8.4 and its checksum. No manifest/package version, public API or wire-format change.Fixed AAC source proof: The quiet source batches 16 AAC frames/371.52 ms;
-pes_payload_size 0still permits 13/about-302 ms, while-max_delay 0gives one frame/PES. Packed and paced sources have byte-identical AAC access units. Fixed-250 uses the paced broadcast; adaptive keeps seven-/sixteen-frame packing. Ceilings, timeouts and durations are unchanged. Source pacing makes fixed-250 attainable; exact hosted outcomes appear below.Browser memory: the watch render and publish capture worklets stop before their
AudioContextcloses, so closed audio graphs are collected.Relay CPU: in this PR's first pass, every lite-05+ viewer's TRACK_INFO left a parked segment that made the relay recompute the live edge on nearly every poll (+18% instructions at 4x10). Fixed, and the per-subscriber splice state is now smaller than the original's (584 to 368 bytes).
Unslop: 22 commits, net -959 lines. Test-only files no longer ship in
@moq/watchand@moq/hang.Trace retention
a99d83dbd. Save the already-active trace after actualpage.close()and the existing pagehide beacon flush, before context/browser close. The old AB driver exited 0 and flushed its beacon but lost the ZIP; the fixed driver retains actual DOM snapshots and completedPage.closerecords. CI runs the regression. Playback, budgets and production/package/public API/wire are unchanged; this repairs evidence loss, not audio starvation.Merge with a merge commit, not a squash, so each issue can name its fixing commit.
Impact
Public API, all additive:
moq_net::track::Producer::break_at(sequence)and@moq/netTrack.Producer.breakAt(sequence): a subscription made afterwards starts atsequenceat the earliest; a fetch still reaches the groups before it.@moq/hangContainer.Legacy.writeMarker(group, end): writes the endpoint marker as the whole group and closes it.moq_tokio::Client::close(): closes the client's QUIC connections once each peer was sent the close (fix(cli): close the relay connection on SIGINT and SIGTERM moq-dev/moq#4287). TheNoqClient::close()it calls is crate-private.The takeover CSP documentation also states that the blob audio worklet requires
script-srcto allowblob:, including the main-thread audio path.Wire: no format change. One reset code moves (#46):
0x34) on moq-lite, like Rust, instead of DELIVERY_TIMEOUT (0x2). IETF drafts have no Old, so there it now sends INTERNAL.@moq/netExpiredcarries0x34, a peer's0x34decodes to it, and a peer's0x2stays a plain stream error. The moq-lite draft says so.Behaviour:
moq-muxcontainer::Producer::discontinuitydeclares the track's break, so a late joiner starts at the marker.@moq/hangContainer.Consumer: a stale group dropped behind returned media counts inskippedbut raises no discontinuity, so the live GOP is not cut.moq-shaper --report-interval 0sis a usage error.@moq/watchand@moq/hangno longer ship test-only files.Historical relay CPU and RAM against the original
1217ea74b:The VPS relay rows in the table measured
0df56d52a, before the final catalog correction. The Mac and Criterion rows ran the relay branch head23354da3d, before the publisher break and the finished-cache confirmation correction. The 2 to 3 segment cost exists only while a splice holds several live routes (a takeover); it is the price of moq-dev#4103's drift anchor. Full tables go in the results comment on #3.Alternatives
Follow-ups
ffmpeg ... -f mpegts -pes_payload_size 0 - | moq import tsline holds quiet audio back for up to 0.35 s #54 is unchanged (upstream docs and demo).AudioContext; a relay already subscribed when the publisher declares a break does not learn it; about 50 players per Chromium, its ownAudioContextlimit; departing viewers hold a separate warm cache for 30 s, which raises relay heap on both builds. The warm-cache RAM cost is a separate task.1217ea74band actualdf3d2b444were grader OK: Chromium had 1/1 underruns and 780/760 ms concealed (frame gaps 602/601.8 ms), WebKit 2/2 and 200/220 ms, Firefox 0/0. No improvement or regression is attributed to the 20 ms differences. Both Chromium source-PTS windows include quiet at 462.6 s; publisher elapsed time is unknown without a logged connection. These supplement the old 0df matrix, not a full final-head matrix.Actual-head long rows, original /
df3d2b444, each 480 s:All six raw graders are OK; these retain the nonzero events and low FPS. Source windows differ across browsers; both Chromium windows include the known quiet passage. These are quality observations, without CPU performance attribution. Earlier 0df rows remain separately attributed.
Linux completed 40 rows, a coverage count: one guarded tiles cell interrupted, 11 absent and two older catalog joins failed. The real-Origin regression fixes their shared-cache mechanism; seven once-only
a3f1a74f8rows retained five grader OK and two sampler failures, detailed below. Firefox's 800 ms quiet stretch is valid but source-phase confounded: original missed that phase, and matched original/final replay has identical underrun, skip and concealment counters. It cannot be attributed to this PR.bc2680217, three Playwright-owned real-task fixtures passed: 0/23 healthy burst heads flagged with and without the probe; actual 150 ms freezes and 660 ms busy tasks flagged. Stall source is byte-identical ata3f1a74f8anddf3d2b444. Six playback measurements retain five oversize-metadata provenance voids and one compact-metadata WebKit main pass, with a later underrun/760 ms concealment. Chromium had 20 hidden samples per row, Firefox/WebKit none; their tab actions did not prove hiding. No void was retried/regraded. These bc268 diagnostics verify neither the catalog fix nor real Safari.df3d2b444. Prior31/38 coordinator rows:18 OK/four strict-CSP voids/nine problems; prior32 raws one partial/six untouched, V5raw33/two partials/five untouched. V6 ran those five once:four OK/one problem. Final36 coordinator rows:22 OK/four CSP voids/ten problems;38 raws include two interrupted partials, none untouched. Six worker-only CSP PASS; four script-header refusals,two/build. Six restarts FAIL suspended/time0 despite input/video; resume/activation cause unproved. Both DF memory rows fail all10 recoveries like original r1/r2; contexts/nodes10/1 are not heap proof. DFtilesweep1:three players/expected one, BSDseq 1 0emitted1,0. DF10 partial SIGTERM:painted0,sampler1,001>1,000ms,missing second input read. V5DF50 guard SIGTERM/driver VOID: designated paint6.9median/4.7minfps,video-input FPS24.1median;5.4s,all50 readiness unknown,user severe lag/possible regression. Before V6 original50 never ran; V6original50 FAIL:paint5.6/0fps,gap6745ms,skew5537.5ms,5underruns/4460ms concealment;45s versus partial5.4s cannot grade a regression. Original/DF480s long OK each1underrun/560ms concealment,gap616ms,paint minima13.5/9.5fps. Raw decoded/decodeFps observes compressed frames before submission, not proof of submission/output; paint measures one designated renderer. Prior lock/V4 refusal retained; no retry/regrade. Checkpoint75/17cells/disjoint21names retained. Other2,713 objects/Cargo.lock matchab075bbb0; metadata DF. AAC13-/16-frame source explanation/receiver ambiguity retained; WebKit separate.V6 original1/10 passed:designatedpaint24.1fps median,min23.1/22.8,gap67ms,skewp9539.2/33.5ms,zero underruns/concealment;all1/10 input snapshots complete. Original50 failed:paint5.6/0fps,video-input FPS17.8/0,gap6745ms,skew5537.5ms,5underruns/4460ms concealment,13video stalls/3153ms/four rebuilds. All50 input reads complete,28 meet raw input-rate threshold/22 flags including two null,physicalonScreen18; output readiness unknown. Audioadvance40402.2/45000ms,PCMoutput45322.4ms. Original/DF480s long both graderOK:each1under/560ms concealment/gap616ms,paint24median/min13.5/9.5,skew40.8/39.3ms;audioadvance464703/464677 and464736.3/464764ms,PCMoutput465008.6/465023.1ms,two audio-stalled samples each. No recorded listening/amplitude evidence or equivalent completeDF50 window; setup/clock progress is not subjective sound-quality proof. All raw decoded/decodedFps/perTile.decodeFps/decoding fields observe compressed frames before submission (both1217 andDF). A closed-decoder break precedes decode, so these counters prove neither successful submission nor decoded output. Painted counts one designated renderer timestamp,not all50. Raw fields/grades stay unchanged. V6guard clear/no retries; all recorded owned processes and private ports absent after teardown.
The user reported, in consecutive messages, "I noticed video is laggin very very hard", "This feel like a regression", then "The Safari test with many players". The V5 partial retains that adverse experience; it is neither a pass nor proof of a source regression. Metrics refer to selected
bbb.hang, not all50: painted6.9fps median/4.7 minimum, video-input FPS24.1 median/22.9 minimum, skewp95=143.9ms; all five measured seconds below15fps. Fifty DOM players/visible-always gates existed, but perTile/decoding/onScreen are null because the final all-tile read failed. Only5.4s/20 samples followed15s warmup; frame log37 paints/243ms maxgap, sampler806ms maxgap, two rebuilds for5s/10s no frame. Twelve CPU samples span23.1s; browser-tree RSS peaks1,391.8MB/system load1 peaks21.97, excluding publisher RSS. AudioContext advances5,364.4ms/5,388ms wall and PCM-output counters5,674.4ms; zero sampled underrun/concealment is not a valid quality pass for an interrupted driver-void. No original50 comparator had run at V5; V6 later measured it once, with a different45s versus5.4s partial window. All recorded owned Safari/WebKit/50publisher/relay/page-server/caffeinate PIDs and private ports were gone in the retained teardown. Those five untouched cells completed once in V6; actualDF50 stayed excluded from repetition. At V5, shipping PR57 was19b/Interop red; separate8509 diagnostic was not a shipping-head replacement.bc2680217Audio quality passed 172 replay checks/14 rows, 36 enforced browser checks/24 measured rows and a separate 24-check fixed-250 isolated main-thread control/two rows. It also reports eight non-enforced metric breaches/six rows and three non-enforced playout voids; two raw silence exceedances pass complete quiet-source proof. A separate bc268 Nightly fixed-250 shared/main run failed: 12 enforced AAC metric breaches in its full matrix; its separate shared/main control had Opus 6.5 underruns/min, 0.032 stalled share, 6.5 skips/min and 905.7 skipped ms/min (four enforced breaches), plus one AAC playout void. Manual success is not a universal pass.The
a3f1a74f8Nightly Audio quality job (run 36819237894) passed 172 replay checks/14 rows, 36 enforced worker checks/24 rows and 24 shared/main checks/two rows. It retained 11 non-enforced metric breaches/seven rows and four non-enforced playout voids; one worker Opus and one main AAC raw silence exceedance passed complete quiet-source proof. Its main Nightly job failed the yankedyoke-derive0.8.3 audit, addressed by the later lock update.At
df3d2b444, manual Audio quality 36822658662 and Nightly Audio quality job 110241399977 both passed 172 replay checks/14 rows, 36 enforced worker checks/24 rows and 24 shared/main checks/two rows. Excluded diagnostics remain: manual 12 metric breaches/eight rows and one playout void; Nightly five breaches/four rows and five voids. Complete quiet-source proof accepted one/two worker raw silence exceedances respectively and one main AAC exceedance in each run. Overall Nightly 36822661339 failed the macOS auth publisher-connect test; only its Audio quality job passed.At actual
2a78f7425, manual AQ 36831617235 succeeded: 172 replay checks/14 rows, 36 worker/24 rows, 24 shared/main/two rows. Worker exclusions: 23 metric breaches/seven rows, five playout voids; three worker/one main silence exceedances pass complete quiet-source proof. Nightly AQ job 110269314921 failed eight enforced fixed-250 main-control metrics, with zero main voids; replay172/14 and worker36/24 passed. Worker exclusions: 17 breaches/six rows, six voids; three worker/one main quiet-source passes. Both main codecs had 1.1 underruns/min, 0.009 stalled share and 1.1 skips/min; skipped ms/min were AAC90.6, Opus370.9, all against zero ceilings. Overall Nightly 36831620398 failed: 75 successful jobs, one AQ failure, 15 skipped. Failure owner is unknown; production source equality with DF/AB does not relabel this as an AB measurement.At actual
ab075bbb0, manual AQ 36835459388 succeeded: 172 replay checks/14 rows, 36 worker/24 rows and 24 main/two rows. Worker exclusions: 11 metric breaches/five rows and one playout void; six silence exceedances pass complete quiet-source proof. Main has zero breaches/voids and one quiet-source accept. These remain AB measurements, not a99 results.Actual AB Interop 36835086137/job 110280467248 failed: 24 pairs passed; all eight Python-publisher pairs failed at 30 s. Python connected, announced catalog and produced audio; browser audioBytes19514/audioTS16860 advanced with running worker audio, but videoFrames0/paintedfalse. Native video reads reset Old
0x34at frame0 and later groups throughout 30 s. Cause is unresolved. All 32 local AB pairs and earlier hosted DF Interop passed; Python source matches DF.The source/artifact census found no recurring healthy five-second actor in the reviewed app, AQ harness or Playwright code. The starvation initiator remains unknown. Worker/main comparisons differ in feeding thread/session, ring backend and source phase, so a worker PASS does not establish causation. Local movie timing does not prove CI source timing.
ac6461ae2changes onlya_stage_broadcast_picks_the_catalog_to_read: full parser/real TCP/Origin selects literal["hd"]/["stereo"]at elapsed ZERO (.070 s); both sessions/drivers/server join before clock release. Production BUDGET500/CEILING1500/SETTLE30 ms and other TLS/QUIC tests stay unchanged. A99 broad red: 838 pass/one fail/three skip, 4,127 not run,[]versus["stereo"]; focused original .103 s PASS, paused-only .099 s PASS without stasis proof, frozen QUIC60.008 s TIMEOUT discarded. Local check/Interop passed; completed AC hosted outcomes appear below. No CPU or product-cause claim.Actual AB Nightly 36835462958 succeeded: 76 jobs successful/15 skipped. Its AQjob 110281696946 passed 172 replay checks/14 rows, 36 worker/24 rows, 24 main/two rows; worker exclusions10 metric breaches/four rows and four voids, four complete quiet-source accepts; main zero breaches/voids, one quiet-source accept. This remains AB, not AC evidence; AB PR Interop failed separately.
The interrupted VPS diagnostic retained five 540.8–568.7 ms page gaps with zero sampled underruns/concealment. Explicit executed-block markers were lost; Opus controls never ran. The unchanged pressure guard stopped the attempt, so it has no completed quality verdict and does not explain the earlier Nightly starvation.
main(CONFLICTING/DIRTY) at its pre-merge head1217ea74b; its old Interop failure remains historical evidence. GitHub does not launchpull_requestworkflows for a conflicting base. GitHub documents this trigger limit. PR feat(watch): add adaptive audio playout and worker offload #3 keeps its existing target. Verification of the merged debug branch is reported separately and does not verify a merge into currentmain.df3d2b444Nix check from0df56d52astays green. Atab075bbb0, Nixjust checkfrom DF passes, including 339 relay tests/two skipped;just test interop --allpasses all 32 pairs across Rust, Python, Go, browser JS, Node, Bun, C and GStreamer.auth_lifetime::an_outage_keeps_the_session_until_expireson macOS: publisher connect timeout; its Audio quality job passed.Actual AC PR Test36846958575/job110319278310 FAIL: 4,769 pass/one fail/six skip/318 unrun; Linux
dropping_a_server_keeps_its_socketcounts one versus two after both drops. Test unchanged from1217ea74b; observer counts network namespace. Actual initiating owner unproved; no CPU/production-retainer attribution. Other AC hosted runs continue once.Actual
ac6461ae2manual AQ36847001121 and Nightly36847004709 PASS; Nightly76 successful/15 skipped jobs. Each AQ: replay172/14, worker36/24, main24/two rows; zero enforced breaches/voids. Worker exclusions, manual/Nightly:14 breaches/six rows/four voids/five quiet accepts; nine/seven/four/three. Each main: zero breaches/voids, one quiet-source accept. Nightly has56 release artifacts, no AQ archive. Full raw rows remain in PR #57; no source-cause or19b result follows.Actual AC Nightly worker raw rows:
Actual AC Nightly sharedMain raw rows:
Actual AC worker raw rows:
Actual AC sharedMain raw rows:
19bcd9d63changes only the Linux worker test, +66/-26 lines. It snapshots process-owned socket inode IDs without opening/duplicating a socket: exactly two after split, the same IDs after each unused member drop, none after shutdown. A plain bind must return AddrInUse while alive and succeed after shutdown. The single-test child isolates parallel libtest siblings; its fresh literal receipt after every assertion prevents a zero-test filter passing. Procfs errors fail except disappeared unrelated FDs; unrelated non-UTF8 file targets are ignored. Docker barrier: foreign same-port exit changes namespace count 3 to 2 while owned IDs stay stable; deliberate owned closure gives 1 then 0 and both bind controls pass. This proves the old observer attribution defect, not the initiating owner of AC's historical 2-to-1 failure. No CPU/production-retainer attribution or production/API/wire change. Actual Linux Test36853755318/job110341245926 passed:5,088 pass/six skipped78.659s; socket test0.052s PASS. Mac skips this fixture. PR Interop remains failed below.Exact AC gates:
ac6461ae2is pushed. Local: 4,966 pass/three skip + 84 pass/zero skip; JS/build/lint; Interop 32/32 PASS. Seven PR checks pass; Test fails above. Manual AQ36847001121 PASS; Nightly36847004709 PASS:76 success/15 skip. Earlier 2a/DF Nightly and AB Interop failures stay retained; AB manual AQ and whole Nightly succeeded.Actual
19bcd9d63PR Interop36853755161/job110341245092 FAIL: all 32 publisher/subscriber pairs and worker media lifecycle40s passed. Main-thread lifecycle failed at14s: after resume,67/81 samples carried the fixture tone above the noise floor. Four negative controls passed: frozen video6s, silent audio6s, offset audio7s, leaked session12s. This differs from AB's eight Python-publisher pair failures. The initiating owner is unresolved; local32-pair PASS remains separate. That19bcd9d63candidate was not approved for merge.19b historical gates:
19bcd9d63is pushed; local 4,966 pass/three skip + 84 pass/zero skip, full JS/build/lint and Interop32/32 PASS. Linux Test36853755318/110341245926 PASS:5,088 pass/6 skip78.659s; socket0.052s PASS; Mac skips it.Candidate gates:
09a783a78is pushed; Local 4,966 passed/three skipped (72.071s) + 84 passed/zero skipped (4.513s); JS/build/lint; 32/32 Interop PASS. Eight actual PR checks SUCCESS.Actual
09a783a78hosted results: Manual 36874501140 SUCCESS (2 jobs: 1 skipped/1 success); replay 172/14 SUCCESS; worker 36/24 SUCCESS, reported 10 metrics/6 rows/3 voids/3 quiet accepts; main 24/2 SUCCESS, reported 0 metrics/0 rows/0 voids/1 quiet accepts. Nightly 36874506339 SUCCESS (91 jobs: 15 skipped/76 success); replay 172/14 SUCCESS; worker 36/24 SUCCESS, reported 11 metrics/9 rows/4 voids/0 quiet accepts; main 24/2 SUCCESS, reported 0 metrics/0 rows/0 voids/1 quiet accepts.09a783a78: three harness files (+105/-2) surface browser-close errors after every close attempt. Swallow mutant:1 PASS/2 FAIL; fix:3 tests/6 assertions PASS, package6 PASS. Prior midwindow owner remains unknown. No playback/API/wire/budget change. DF differs only in12 test/doc/harness objects;2,705 others/eight AQ inputs match. Binaries stay DF; no compiled-binary/CPU equivalence.Actual
19bcd9d63hosted results: Manual36853751819 PASS: replay172/14, worker36/24, main24/2; worker11 excluded breaches/6 rows/4 voids/no quiet allowance. Nightly36853756326 FAIL:75 success/1 AQ fail/15 skip; replay172/14, main24/2 PASS; worker3 enforced fixed250/plain AAC breaches:1.1 underruns/min,1.1 episodes/min,220.5 underrun ms/min, zero ceilings;26 excluded/10 rows/2 voids. Each main0/0+1 quiet. Nightly timing stays outside merge gates; 19b Interop failed. No failure-owner attribution follows. The worker failure is not36 passed checks. The failure artifact retains24 traces with completed page.close, including both void rows; trace preservation does not make their playback valid.Actual19b manualAQ worker raw rows:
Actual19b manualAQ sharedMain raw rows:
Actual19b manualAQ recorded-only exclusions:
Actual19b Nightly worker raw rows:
Actual19b Nightly sharedMain raw rows:
Actual19b Nightly recorded-only exclusions:
chromium-aac-44100-bursty-plain silence_share_share: 0.25 > 0.075; 53/55 quiet windows over quiet source (log RMS r 1.00000, level 1.000, 65 audible windows); 2 of 55 quiet windows unproven, first at 42.75 s: audible source
chromium-aac-44100-high-rtt-isolated skipped_samples_per_min: 133.5 > 85.9
chromium-aac-44100-high-rtt-isolated target_ms_p95: 440 > 330
chromium-aac-44100-high-rtt-isolated converge_s_seconds: 57.7 > 11.6
chromium-aac-44100-high-rtt-plain underruns_per_min: 2.2 > 0
chromium-aac-44100-high-rtt-plain underrun_episodes_per_min: 2.2 > 0
chromium-aac-44100-high-rtt-plain underrun_samples_per_min: 476.6 > 0
chromium-aac-44100-mild-isolated underruns_per_min: 2.2 > 1.7
chromium-aac-44100-mild-isolated underrun_episodes_per_min: 2.2 > 0
chromium-aac-44100-mild-isolated underrun_samples_per_min: 376.4 > 0
chromium-aac-44100-near-zero-isolated underruns_per_min: 2.2 > 0
chromium-aac-44100-near-zero-isolated underrun_episodes_per_min: 2.2 > 0
chromium-aac-44100-near-zero-isolated underrun_samples_per_min: 469.9 > 0
chromium-aac-44100-near-zero-plain underruns_per_min: 2.2 > 0
chromium-aac-44100-near-zero-plain underrun_episodes_per_min: 2.2 > 0
chromium-aac-44100-near-zero-plain underrun_samples_per_min: 478.5 > 0
chromium-aac-44100-near-zero-plain converge_s_seconds: 50.3 > 3
chromium-aac-44100-near-zero-plain silence_share_share: 0.344 > 0.075; 74/76 quiet windows over quiet source (log RMS r 0.99994, level 1.000, 110 audible windows); 2 of 76 quiet windows unproven, first at 27.27 s: discontinuity
chromium-aac-44100-step-isolated silence_share_share: 0.345 > 0.061; 73/76 quiet windows over quiet source (log RMS r 0.99995, level 1.000, 80 audible windows, 2 analyser lag segments); 3 of 76 quiet windows unproven, first at 23.25 s: audible source
chromium-opus-48000-bursty-isolated underrun_episodes_per_min: 1.1 > 0
chromium-opus-48000-bursty-isolated underrun_samples_per_min: 272.7 > 0
chromium-opus-48000-bursty-isolated silence_share_share: 0.177 > 0.15; 36/39 quiet windows over quiet source (log RMS r 0.99995, level 1.000, 117 audible windows); 3 of 39 quiet windows unproven, first at 39.50 s: audible source
chromium-opus-48000-high-rtt-isolated silence_share_share: 0.045 > 0.007; 8/10 quiet windows over quiet source (log RMS r 0.99923, level 1.000, 145 audible windows); 2 of 10 quiet windows unproven, first at 9.60 s: discontinuity
chromium-opus-48000-step-isolated skip_aheads_per_min: 5.5 > 4.9
chromium-opus-48000-step-isolated skipped_samples_per_min: 2902.2 > 419.4
chromium-opus-48000-step-isolated silence_share_share: 0.23 > 0.088; 45/48 quiet windows over quiet source (log RMS r 1.00000, level 1.000, 81 audible windows); 3 of 48 quiet windows unproven, first at 21.16 s: discontinuity
chromium-aac-44100-bursty-isolated: playout (the audio graph, timeline, or counters changed during the measured window)
chromium-opus-48000-high-rtt-plain: clock (AudioContext.currentTime drifted 3.06% from wall clock over 10s)
Catalog browser follow-up at
a3f1a74f8: Seven once-only Linux rows useda3f1a74f8: five grader OK; two retained sampler failures. Chromium reattach r1 had a 2,078.6 ms gap after attach, one whole-window underrun/1,160 ms concealed; r2 had a 1,748.8 ms gap in warmup, zero underruns/concealment. Other whole-window counts were r3 reattach 2/120 ms, r3 hidden 1/120 ms, Firefox hidden 2/40 ms and r1/r2 hidden 0/0. Matched r2/r3 hidden catalog-to-media was 61.2/64.4 ms and click-to-ready 145/179 ms; Firefox was 22/1,341 ms. NoLaggedor no-route warning appears in the relay log. These warm joins do not force the internal two-front overlap; the native real-Origin test does. Production Rust/JS matchesdf3d2b444; a3 used the prior lock.ab075bbb0adds only auth-test changes;ac6461ae2adds only trace/docs and completion-test changes; production remains unchanged. These native binaries remain attributed toa3f1a74f8, with its prior dependency lock.Drafted by Claude Opus 5.5. Edited and posted by Codex.
(Written by GPT-6)