diff --git a/js/watch/src/audio/buffer.test.ts b/js/watch/src/audio/buffer.test.ts index b55df65179..0c299f4b1c 100644 --- a/js/watch/src/audio/buffer.test.ts +++ b/js/watch/src/audio/buffer.test.ts @@ -443,7 +443,7 @@ describe("Backpressure", () => { it("counts a frame still decoding where its output lands", () => { const gate = new Backpressure(1000, true, Duration.millis(2000)); gate.observe(ms(0), false); - gate.decoded(ms(520), ms(500)); + gate.decoded(ms(520), ms(20)); gate.inserted(ms(480), ms(500)); const signal = new AbortController().signal; expect(Bun.peek.status(gate.wait(ms(520), signal))).toBe("fulfilled"); @@ -580,7 +580,7 @@ describe("AudioBuffer steering", () => { const { buffer, sent } = fake(100); try { buffer.received(Duration.millis(1000)); - buffer.decoded(ms(1020), ms(1014)); + buffer.decoded(ms(1020), ms(6)); buffer.received(Duration.millis(1040)); expect(sent.slice(1)).toEqual([ { type: "received", end: 1_000_000 }, @@ -831,4 +831,29 @@ describe("AudioBuffer steering", () => { clock.close(); } }); + + // Opus output lands on the ring a pre-skip before its frame timestamp, at its presentation time, so + // the clock maps onto the ring unshifted and the playhead reports the ring's position: only a frame + // timestamp, as `received` takes, moves back by the pre-skip. + it("plays and reports the ring at presentation time under a pre-skip", () => { + const clock = started(); + const { buffer, sent, state, context } = fake(100, { + clock: { position: () => 5000, out: { reference: new Signal(0) } }, + }); + try { + buffer.decoded(ms(5020), ms(6.5)); + buffer.received(Duration.millis(5020)); + context.currentTime = 1; + context.sounding = 1; + state(ms(5000), false, 0, 1); + expect(sent.filter((msg) => msg.type === "clock")).toEqual([ + { type: "clock", timestamp: 5_000_000, time: 1 }, + ]); + expect(sent).toContainEqual({ type: "received", end: 5_013_500 }); + expect(buffer.timestamp.peek()).toBe(ms(5000)); + } finally { + buffer.close(); + clock.close(); + } + }); }); diff --git a/js/watch/src/audio/buffer.ts b/js/watch/src/audio/buffer.ts index 59d657e820..440d49a494 100644 --- a/js/watch/src/audio/buffer.ts +++ b/js/watch/src/audio/buffer.ts @@ -145,12 +145,12 @@ export class Backpressure { } /** - * Note that the decoder produced output ending at `end`, which lands in the ring up to `landed` - * unless it was discarded. It answers the oldest frame still decoding and every one it reaches - * past: a decoder may emit nothing for a frame. + * Note that the decoder produced output ending at `end`, which lands in the ring `offset` behind it + * (the codec's pre-skip, not what a trim removed) unless it was discarded. It answers the oldest + * frame still decoding and every one it reaches past: a decoder may emit nothing for a frame. */ - decoded(end: number, landed?: number): void { - if (landed !== undefined) this.#offset = end - landed; + decoded(end: number, offset?: number): void { + if (offset !== undefined) this.#offset = offset; let answered = 0; while (answered < this.#decoding.length && (answered === 0 || this.#decoding[answered] < end)) answered++; this.#decoding.splice(0, answered); @@ -295,7 +295,7 @@ export class AudioBuffer { #context: AudioContext; readonly #timestamp = new Signal(0); - /** The playback timestamp sounding now, as of the worklet's last report. */ + /** The presentation timestamp sounding now, as of the worklet's last report. */ readonly timestamp: Getter = this.#timestamp; readonly #output = new Signal(Duration.ZERO); @@ -520,7 +520,9 @@ export class AudioBuffer { } // A reference move is a step of the clock, not where a reading landed, so the offset takes it whole. const moved = last !== undefined && reference !== undefined ? (last - reference) * 1000 : 0; - const offset = position * 1000 - this.#backpressure.offset - heard * 1_000_000; + // The ring already holds presentation time (Opus output lands a pre-skip before its frame + // timestamp), so the clock maps onto it unshifted. + const offset = position * 1000 - heard * 1_000_000; const previous = this.#offset === undefined ? undefined : this.#offset + moved; this.#offset = previous !== undefined && Math.abs(offset - previous) < JUMP ? Math.max(offset, previous - FALL) : offset; @@ -538,10 +540,10 @@ export class AudioBuffer { /** * Note that the decoder produced output ending at `end` on its own timeline, which lands in the ring - * up to `landed` unless it was discarded. + * `offset` behind it unless it was discarded. */ - decoded(end: number, landed?: number): void { - this.#backpressure.decoded(end, landed); + decoded(end: number, offset?: number): void { + this.#backpressure.decoded(end, offset); } /** Note that the decoder closed or reset, dropping every frame still decoding. */ diff --git a/js/watch/src/audio/decoder.ts b/js/watch/src/audio/decoder.ts index 30041a8cb2..7f5ef91180 100644 --- a/js/watch/src/audio/decoder.ts +++ b/js/watch/src/audio/decoder.ts @@ -1091,7 +1091,7 @@ export class Decoder { try { landed = this.#place(sample); } finally { - if (!preroll) this.#ring.peek()?.decoded(decoded, landed); + if (!preroll) this.#ring.peek()?.decoded(decoded, landed === undefined ? undefined : this.#terminal.delay); } return landed; } diff --git a/js/watch/src/audio/replay.ts b/js/watch/src/audio/replay.ts index b6b484f923..772dec76c3 100644 --- a/js/watch/src/audio/replay.ts +++ b/js/watch/src/audio/replay.ts @@ -325,7 +325,7 @@ export async function* replay(trace: Arrival[], options: Options): AsyncGenerato const end = timestamp + micros((pcm.length / rate) * 1000); gate.inserted(timestamp, end); decoded.update((span) => ({ start: span?.start ?? media, end: Duration.micros(end) })); - gate.decoded(end, end); + gate.decoded(end, 0); } })().catch((error: unknown) => { failure = { error }; diff --git a/quest/m0/nightly/audio-quality-p50.md b/quest/m0/nightly/audio-quality-p50.md deleted file mode 100644 index a9cde75e63..0000000000 --- a/quest/m0/nightly/audio-quality-p50.md +++ /dev/null @@ -1,17 +0,0 @@ -# [S] Browser audio-quality lag ceilings match what the player corrects - -## Goal - -The Nightly `Audio quality` job's browser `av_lag_ms_p50` ceilings sit outside the lag the player deliberately leaves alone, so a row fails only when audio really drifts. It failed on 11-13, 11-14, 11-15 and 11-17 on `chromium-aac-44100-congestion: av_lag_ms_p50 2.7 > 1.3`. - -## Plan - -Cause: the 1.3 ms ceiling was set provisionally in [moq-dev/moq#8459](https://github.com/moq-dev/moq/pull/8459) from a few developer-machine runs. [moq-dev/moq#8722](https://github.com/moq-dev/moq/pull/8722) then made the player leave lag within ±5 ms (half a 10 ms stretch step) uncorrected, so where p50 settles inside that window is chance (runner: 0.4–18.2 ms across 11-13..11-17). #8722 re-measured the replay rows but left the browser rows to the runner. - -Decided 2026-11-18 (maintainer, /quest-plan): ✅ set the browser p50 ceilings to the ±5 ms window now, then tighten from runner artifacts (`gh run download -n audio-quality-run`) once a week of nights since #8722 exists (rejected: wait for a week of runner data first and apply `budgets.json`'s worst night + 25% + 20 ms rule, about 24 ms, leaving Nightly red meanwhile). - -Also look at the Opus browser rows: several rest 7–11 ms behind on the runner, outside the ±5 ms window, and Opus congestion hit exactly its 8.3 ceiling on 11-17. Decide from the data whether that is a playout bug for [audio playout](/quest/m2/audio-playout/README.md) or a ceiling to set. - -A developer machine can't reproduce the runner's numbers (every browser row sits about 30 ms behind locally, as `budgets.json` notes), so use runner artifacts. - -Public API: none. Wire: none. diff --git a/quest/m0/test-flakes-3/announce-consumer-close.md b/quest/m0/test-flakes-3/announce-consumer-close.md new file mode 100644 index 0000000000..6579f4193d --- /dev/null +++ b/quest/m0/test-flakes-3/announce-consumer-close.md @@ -0,0 +1,20 @@ +# [S] Closing a JS announce consumer ends the peer's stream under load + +## Goal + +`moq-lite-07-wip: closing an announce consumer ends the peer's announce stream` +in `js/net/src/integration.test.ts` passes under a loaded `just check`. + +## Plan + +Seen once on 2026-11-19 while validating +[moq-dev/moq#9139](https://github.com/moq-dev/moq/pull/9139), which touches no +`js/net` code: `until(() => ended === i)` hit its 1 s wall-clock deadline +during the full `test js` run, and the test passed alone right after. + +Find whether the peer's announce stream really ends late under load (the +close reaching the server only on a later turn) or whether the 1 s `until` +deadline is what fails. Then assert on the stream's end as an event instead of +polling against wall time. + +Public API: none. Wire: none. diff --git a/test/audio-quality/budgets.json b/test/audio-quality/budgets.json index 19fdb4937c..07d578430b 100644 --- a/test/audio-quality/budgets.json +++ b/test/audio-quality/budgets.json @@ -1,5 +1,5 @@ { - "note": "The browser rows are measured on the nightly runner (ubuntu-24.04-arm, headless Chromium, seed 7, 60 s a row), from the audio-quality-run artifacts of the seven scheduled Nightly runs on main from 2026-10-29 to 2026-11-04 (40885347517, 41065371094, 41204380716, 41313466339, 41483654426, 41656613559, 41852517544), the first after a fresh subscriber stopped receiving the relay's cached history. The six AAC rows other than congestion are measured the same way from the seven runs from 2026-11-05 to 2026-11-11 (42050683248, 42236107954, 42377867523, 42487375545, 42659882487, 42842827718, 43041909115), the first after the TS importer published each PES as one group; the first four predate the browser's time stretch, and the six rows keep the four provisional ceilings set below from 2026-11-09 and 2026-11-10. Each ceiling is the worst of its seven nights times 1.25, plus one runner hiccup, rounded up to two significant figures. The hiccup is sized from the single gaps the 2026-10-08 night put on otherwise clean rows, the fixed-250 control among them: 250 underruns/min, 700 gap ms/min, 3.3 skips/min, 280 dropped ms/min (both sized on the playhead's skip-ahead inference, and now applied to player_skips_per_min and trimmed_ms_per_min), 0.01 stalled, 0.015 silence, and one 20 ms bucket of target. Budgeted this way, every night passes when it is held out of its row's measurement except one, and that row's own ceiling includes it: the 2026-11-03 Opus bursty row raised its target to 300 ms 2.6 s in and held it for one 20 s half-life before settling at 180. Such startup holds, Opus near-zero's 100 ms on 2026-10-30 and 10-31 among them, are budgeted as measured. These ceilings record today's player as the runner hears it, not target audio quality. To re-measure, download the same artifact from the latest scheduled nightlies (`gh run download -n audio-quality-run`), which keep every run, and apply the same rule. On 2026-11-10 the inferred skip_aheads and discarded_ms keys were dropped, since they read a time stretch's catch-up as dropped audio, and every browser row gained player_skips_per_min, trimmed_ms_per_min, and av_lag_ms p50 and p95, by the same rule from the only two scheduled nights on main with today's browser player and its trim counter, 2026-11-09 and 2026-11-10 (42659882487, 42842827718). A lag is signed, so its ceiling is the worst night plus a quarter of its size plus one 20 ms bucket. These four keys are provisional until seven nights with the stretch model re-measure every row. The grade follows the model: audio rests on the shared clock, trails it through a stall, and catches up after. When a change moves a row, a steady-state lag (av_lag_ms p50) that rises, or audio the player trimmed, is a regression to fix rather than budget; a p95 lag that rises through a stall, or an underrun count that moves, is re-measured in. No ceiling is fixed at zero underruns or zero stalls; a replay row's is exactly what it measured. The replay rows are deterministic, so their ceilings are exactly what they measured, through the end of each recording, with no headroom: any regression fails them, and an improvement should be tightened in. They were last re-measured on 2026-10-22 using version 3 traces with encoded frame durations: AAC-LC uses its fixed 1024 samples, and relay-mic retains Opus packet TOC durations. relay-mic was re-recorded on 2026-11-03 from a real microphone, Chrome 156 on a MacBook Air publishing its built-in microphone through the public relay, recorded by headless Chromium on a Linux host (35 s, minimum RTT 2 ms), which raised its target ceiling from 60 to 80 ms and its silence from 0 to 0.001 as measured. The recording path itself has no added RTT. relay-mic-rtt40 is the 2026-11-03 Chrome row at 40 ms of added RTT on that laptop (180 s, minimum RTT 47 ms), kept as the baseline time stretch improves on; like every replay row, its player_underruns_per_min ceiling is what it measures, not a fixed zero. player_underruns_per_min, the player's own counter that its stats panel shows, is budgeted on the replay rows only; it counts the run-dry re-stalls the tap files as stalled time. On 2026-11-07, after the ordered read stopped delivering groups below its cursor (#8071) and took Rust's hole and age bounds (#8218), relay-bbb was re-measured at 16.1 player underruns/min, 0.025 stalled and 0.02 silence, and relay-mic-rtt40 at 8.2, its discarded audio at 70.9 ms/min and its target p95 at 320 ms rather than 780, since the stale backlog it never played no longer reaches the estimator; neither row lost audio the previous player heard. On 2026-11-08 the player began time-stretching live playback with NetEq's accelerate, deciding on the ring's depth right after each arrival and never expanding, and every replay row was re-measured with it, the browser rows left to the nightly runner. Each row's av_lag_ms, how far the playhead trails the shared clock video plays from, is what it measured: the ring rests one frame ahead of the newest arrival's playout, as the trim it replaces left it, so audio leads the clock by 19 to 23 ms at p50. Against the trim, the A/V lag p50 fell from 0.7, -17.5, -21.5, -18.2 and -15.6 ms to -22.4 on bursty-aac, -22.2 on local-aac, -22.7 on relay-bbb, -19 on relay-mic and -20.1 on relay-mic-rtt40, and its p95 fell on every row, most on relay-bbb (10.5 to -4.2) and relay-mic-rtt40 (-0.9 to -13.4). Underruns stayed where the trim had them, as the latency they would cost is not spent: player_underruns_per_min held at 4.4, 12.1, 16.1 and 6.1 and rose from 8.2 to 10.6 on relay-mic-rtt40, and relay-mic's tap now files 2 short quanta a minute, 2 ms of missing audio, where it filed none. What the stretch buys is the audio the trim dropped: discarded_ms_per_min fell from 87.8 to 0 on relay-bbb and from 127.5 to 100.6 on local-aac, whose remainder drops nothing: two receive stalls 14 and 15.3 s in leave audio 159 and 132 ms behind the clock, and the fast accelerate brings it back within half a second. player_skips_per_min and trimmed_ms_per_min, the ring's own count of re-anchors that dropped buffered audio and the media they dropped, are budgeted beside it, at 0 on every recorded row. relay-mic-stall, added 2026-11-10, is relay-mic with nothing arriving from 15 s to 16.5 s and the held frames then trickling in at twice real time until 18 s; today's player trims 2152.5 ms/min of it and its target p95 rises to 1540 ms. Once durations became whole nanoseconds, relay-bbb's delay, 200 ms plus one 1024-sample AAC frame, rounds up by a fraction of a sample, so its ring holds one more sample (23 µs) and its A/V lag p50 measures -22.6 rather than -22.7. Neither replay calibration changed the browser or native ceilings. bursty-aac is the 2026-10-18 AAC bursty isolated capture, recorded with the observer subscribed at the decoder's 2 s budget, so it holds every group the player saw. The native rows are provisional: measured on a loaded developer host (x86_64 Linux, the private PulseAudio null sink, seed 7, 60 s a row) on 2026-10-17, three runs of every row from the debug build nightly uses plus three more of the AAC rows from a release build, which played alike. Each ceiling is the worst clean run times 1.25, plus one native hiccup sized from the unimpaired Opus row's worst run (20 underruns/min, 50 gap ms/min, 2.2 skips/min, 110 discarded ms/min, 0.01 stalled, 0.005 silence, one 20 ms bucket of target), rounded up to two significant figures. On 2026-11-10, after native playback began reporting its A/V lag (#8361), the native rows' player_skips_per_min, trimmed_ms_per_min, and av_lag_ms p50 and p95, and every key of the native and browser congestion rows, were measured by the same rules from three runs of each on a loaded developer host (x86_64 Linux, the debug build, seed 7, 60 s a row), with no void; native audio rests 9 to 95 ms behind its clock at p50. These are provisional until the stretch model's nightly re-measure, which also moves the browser congestion rows onto the runner. On the 2026-10-17 host the null sink's clock ran about 150 ppm slow against the monotonic clock, read within 5 ms at p95 per sample, and its output latency sat between 14 and 28 ms. Five of the 36 AAC rows (none of the 18 Opus rows) voided on the clock: the device stopped pulling for about two seconds mid-row, sometimes after a backend error and a stream restart, which is a native playback defect rather than noise, and the ceilings exclude those runs. Re-measure from the nightly runner's artifacts once it has a week of native rows. On 2026-11-11 every player began following one stretch rule on the shared clock: it slows down below its target, speeds up above it, and skips only past its bound, which in the browser is the subscription's max delay. Every replay row was re-measured with it, the browser and native rows left to the nightly runner. Audio now rests on the clock rather than one frame ahead of it, so the A/V lag p50 rose from about -20 ms to 0.4 ms on relay-bbb, 1.1 on relay-mic, 1.2 on relay-mic-rtt40 and 1.8 on local-aac. Its p95 rose on every row, since audio that falls behind is now caught up by speeding up rather than trimmed: the shared clock steps forward as its reference converges through a tune-in backlog, and again when a target falls, and audio follows at the accelerate's pace, which on the replay's constant tone is one 2.5 ms period a window, slower than on speech. bursty-aac rests 5 ms behind its clock, so its p50 of 10.4 ms is mostly its catch-ups, and relay-mic-stall's p50 of 66.1 ms is the catch-up after its target falls from 1540 to 80 ms, which outlasts half the row. In exchange nothing is trimmed: relay-mic-stall's trimmed_ms_per_min fell from 2152.5 to 0, its skips from 6.1 to 0, and its underruns from 24.2 to 0 a minute, and player_underruns_per_min fell from 12.1 to 4 on local-aac, 16.1 to 0 on relay-bbb, 6.1 to 0 on relay-mic, 10.6 to 2.7 on relay-mic-rtt40 and 6.1 to 4 on relay-mic-stall, holding at 4.4 on bursty-aac. The tap now files 3.3 to 6.1 short quanta a minute on the AAC rows, under 0.3 ms of missing audio, where it filed none. To deliver the audio a stall delays, the container consumer now waits and skips at the subscription's max delay, as native already does, rather than at the delay plus lookahead. relay-mic-rtt40's join backlog arrives reordered, so the consumer now waits on its holes and the estimator observes the frames that fill them, up to 775 ms late, and its target p95 returns from 320 to 780 ms, held for one half-life after the join. Later on 2026-11-11 audio began skipping as far as the shared clock's reference moves earlier, once those moves add up to one stretch window and only as far as it holds audio past its target, rather than speeding through them, and every replay row was re-measured. The tune-in backlog's jumps now land before the graded window, so the A/V lag p95 fell from 453.1 to 2.1 ms on relay-bbb (its p50 from 0.4 to -1.3), 743.7 to 1.9 on relay-mic-rtt40, whose p50 fell from 1.2 to 0.4 and its tap's missing audio rose from 0.6 to 0.9 ms a minute, since a jump past what is decoded skips the rest once it lands,, 335.5 to 39.2 on local-aac and 88.6 to 0.3 on relay-mic, and the p50 from 66.1 to 25.3 on relay-mic-stall, with nothing trimmed in any graded window. The tap's underruns fell from 6.1 to 2 a minute on local-aac and 6.1 to 0 on relay-bbb; local-aac's silence rose from 0.001 to 0.002. Later on 2026-11-11 the container consumer began waiting on a missing group only for the delay plus the lookahead, while still delivering groups that arrived until they age past the max delay, and every replay row was re-measured. relay-mic-rtt40's target p95 fell from 780 back to 320 ms, since the holes in its reordered join backlog no longer delay the frames behind them; its tap's missing audio rose from 0.9 to 1.1 ms a minute and its A/V lag p50 and p95 from 0.4 and 1.9 to 0.7 and 2.4 ms. bursty-aac's silence rose from 0.007 to 0.009 and its A/V lag p95 fell from 162.1 to 156.3 ms, and relay-bbb's lag p50 and p95 moved from -1.3 and 2.1 to -1.2 and 2.2 ms. relay-mic-stall did not move. Also on 2026-11-11 the stretch became a player setting, 2 s by default, and the audio subscription's max delay became the delay plus it rather than the 2 s constant, so the container consumer delivers a late group that much longer before it ages out, and every replay row was re-measured: only relay-mic moved, its A/V lag p50 and p95 from 0.3 to 1.6 ms, under the 2.5 ms pitch period the stretch can move it by. Holding the constant max delay with the new bound reproduces the old values exactly. On 2026-11-12 the replays stopped playing a constant and began playing a seeded speech-like signal (syllables gliding a 90 to 220 Hz fundamental, with pauses held above the silence classifier), since the stretch's pitch search found only its shortest period in a constant, and the browser ring stopped zero-filling a hole of up to a millisecond and two samples between frames, as native stopped writing silence for one: the recorded AAC timestamps land up to 2.5 samples either side of a 1024-sample boundary, and on speech the 2-sample holes this filled ended 119.1 quanta a minute short on bursty-aac. Every replay row was re-measured. The tap's underruns fell to 0 on bursty-aac (from 3.3) and local-aac (from 2), and relay-mic-rtt40's missing audio from 1.1 to 0.3 ms a minute while its player underruns rose from 2.7 to 3.1. The A/V lag p95 fell from 156.3 to 104.8 ms on bursty-aac, 39.2 to 15.1 on local-aac, 2.2 to 1.4 on relay-bbb and 501.8 to 484.9 on relay-mic-stall, whose p50 fell from 25.3 to 1.1, and rose from 1.6 to 3.4 on relay-mic and 2.4 to 3.5 on relay-mic-rtt40. Where audio rests moved within one speech pitch period, which is as close as one stretch seam can steer it: the p50 is 11 ms on bursty-aac (from 10.4), -4.2 on relay-bbb (from -1.2), 1.1 on relay-mic (from 1.6), 0.9 on relay-mic-rtt40 (from 0.7) and 1.7 on local-aac (from 1.8). On 2026-11-13 the stretch default fell from 2 s to 200 ms and audio trailing the clock past it began skipping onto the clock rather than back to the bound, as video skips, with the audio subscription still reaching 2 s past the delay; every replay row was re-measured. A target fall past 200 ms now costs one skip rather than seconds of catch-up, and so does a trail past it, and a rise past it pads silence until the clock arrives rather than slowing down: relay-mic-stall skips the 1.46 s it trails after the outage and its 1540 to 1020 and 1020 to 80 ms falls, and pads the 0.5 s its 60 to 1540 ms rise leaves it ahead, where it slowed down for 4 s, so it trims 4678.2 ms a minute where it trimmed none and its stalled share rose from 0.05 to 0.067 and silence from 0.049 to 0.085. The frames it holds through the outage then trickle in at twice real time, so its low buffer slows it to 154 ms behind the clock, where it had sat ahead of it: its A/V lag p95 fell from 484.9 to 118.6 ms and its p50 rose from 1.1 to 2.2. relay-mic-rtt40 skips its 320 to 80 ms fall, 0.3 skips and 83.3 trimmed ms a minute, its p95 from 3.5 to 2.9 ms. bursty-aac skips a 225 ms trail 6 s in, 1.1 skips and 312.7 trimmed ms a minute, its p95 from 102.7 to 49.2 ms and p50 from 11 to 10. relay-mic trims nothing, but its tune-in's skips land on the clock rather than at the bound, which moves where it rests within a pitch period: its p50 from 1.1 to 2.2 ms and p95 from 3.4 to 4.5. Trimming at a target fall or a trail past 200 ms is the rule, not a regression. Also on 2026-11-13, for the same change, the congestion rows' trimmed_ms_per_min ceilings were raised by the same rule from two runs of each on a loaded developer host (x86_64 Linux, the debug build, seed 7, 60 s a row): to 3800 on chromium AAC (worst 2740.3 ms/min), 3700 on chromium Opus (2697.2), 3900 on native AAC (2996.9) and 4600 on native Opus (3532.3), whose player_skips_per_min rose to 15 (9.8). The trim is the outage's audio skipped onto the clock rather than played seconds late: against main on the same host, chromium's A/V lag p95 fell from 1396.8 to 61.6 ms on AAC and 1424.8 to 64.3 on Opus, and native's from 1246.3 to 116.7 and 826.9 to 117.5. These four ceilings are provisional until the nightly re-measures the congestion rows. Also on 2026-11-13, the estimator began counting one outage, the intervals its stalled frames span and the empty ones before them, as one observation, and every replay row was re-measured on top of the snap: only relay-mic-stall moved. Its single stall no longer raises the target, so its target p95 fell from 1540 to 80 ms and the trail the outage leaves is skipped onto the clock at once rather than through the target's falls: 2 skips a minute where it took 6.1, 5122.3 trimmed ms a minute where it trimmed 4678.2, and its A/V lag p95 from 118.6 to 19.7 ms. With 80 ms of buffer rather than 1540, more of the outage plays as silence: its stalled share rose from 0.067 to 0.092 and silence from 0.085 to 0.092. On 2026-11-14 every player began following the clock by its lag alone: a continuous WSOLA time stretch for any content, at a rate proportional to the lag and capped at 10%, which leaves a lag within half a 10 ms hop alone, and slows down while the audio received ahead of the playhead is about to run out, replacing NetEq's buffer-level decision and pitch-period splices. Every replay row was re-measured. Audio now rests within half a hop of the clock rather than within a pitch period: the A/V lag p50 fell from 10 to 0.2 ms on bursty-aac, whose p95 fell from 49.2 to 0.2, from 1.7 to -1.3 on local-aac, 2.2 to 1.8 on relay-mic and 0.9 to 0.6 on relay-mic-rtt40, and moved from -2 to 2.7 on relay-bbb and 2.2 to 3.1 on relay-mic-stall, each within the half hop the rule leaves alone. The p95 rose from 3.6 to 6.2 ms on relay-bbb, 4.5 to 6.9 on relay-mic, 2.9 to 6.9 on relay-mic-rtt40 and 19.7 to 32.1 on relay-mic-stall, and fell from 15.1 to 14.3 on local-aac: a lag under a hop is corrected only once its pace adds up to one. player_underruns_per_min rose from 4.4 to 6.6 on bursty-aac and 4 to 8.1 on local-aac, whose recorded outages outlast what slowing down at 10% bridges, and fell from 3.1 to 1 on relay-mic-rtt40, whose tap's underruns fell from 0.7 to 0.3 a minute; bursty-aac's silence rose from 0.007 to 0.008. Nothing more is trimmed on any row. The browser and native rows were measured on a loaded developer host (x86_64 Linux, the debug build, seed 7, 60 s a row) against main the same day and left to the nightly runner: every browser row's A/V lag p50 now sits between 27 and 39 ms, where the fixed-250 control does on that host, where wide and bursty sat up to 58 ms behind, and every native row's within 15 ms of the clock. Also on 2026-11-14 the estimator stopped folding the intervals a subscription's join spans, those holding a frame due before its backlog finished arriving, and every replay row was re-measured: only relay-mic and relay-mic-stall moved. Their first frame arrives slower than the path after it, so their first interval folds nothing and the target leaves the prior 500 ms later, which moves where audio rests within the half hop the rule leaves alone: relay-mic's A/V lag p50 from 1.8 to -0.3 ms and p95 from 6.9 to 9, and relay-mic-stall's p50 from 3.1 to 1.7. relay-mic-rtt40 did not move: the browser consumer closes its join's stale groups below the cursor unread, so the estimator never sees when that backlog ended, and its target p95 held at 320 ms. Later on 2026-11-14 both container consumers began reporting when each group they close unread arrived, if it is at or below the publisher's largest group as the subscription began, which extends a join still open. On 2026-11-15 the traces gained that largest group (version 4), so the replay's consumer learns it as a wire subscriber does, and relay-mic-rtt40 was re-measured: its target p95 fell from 320 to 120 ms, underruns per minute rose from 0.3 to 0.7, player skips per minute fell from 0.3 to 0, trimmed ms per minute from 83 to 0, A/V lag p50 rose from 0.6 to 2.3 ms and p95 from 6.9 to 8.6, the trade of reporting the join's backlog. No other replay row moved. On 2026-11-17 the browser player began reading audio through the ordered reader, which waits for a hole as long as the ring could still play it and reports each half-open range of keys it gives up on as one gap, and the replay dropped the trace's largest group: a second subscription now observes every arrival for the estimate, the join's backlog included. Every replay row was re-measured. bursty-aac's player underruns per minute fell from 6.6 to 4.4 and silence share from 0.008 to 0.007; audio rests at 0.1 ms until a stall at 19.5 s, from which the stretch catches up to 4.2 ms behind the clock, inside the half hop it leaves alone, so its A/V lag p50 and p95 rose from 0.2 to 4.2 ms. relay-bbb's A/V lag p50 fell from 2.7 to 1.4 ms and p95 rose from 6.2 to 7.6. relay-mic-rtt40's underruns per minute fell from 0.7 to 0.3, as did its player underruns per minute, underrun ms per minute from 0.8 to 0.2, A/V lag p50 from 2.3 to 0.6 ms, and p95 rose from 8.6 to 9.7, its target p95 holding at 120 ms; 0.6 ms of that rise is the replay's ordered reader taking the catalog's warmup into its budget, as the browser's does. No other replay row moved.", + "note": "The browser rows are measured on the nightly runner (ubuntu-24.04-arm, headless Chromium, seed 7, 60 s a row), from the audio-quality-run artifacts of the seven scheduled Nightly runs on main from 2026-10-29 to 2026-11-04 (40885347517, 41065371094, 41204380716, 41313466339, 41483654426, 41656613559, 41852517544), the first after a fresh subscriber stopped receiving the relay's cached history. The six AAC rows other than congestion are measured the same way from the seven runs from 2026-11-05 to 2026-11-11 (42050683248, 42236107954, 42377867523, 42487375545, 42659882487, 42842827718, 43041909115), the first after the TS importer published each PES as one group; the first four predate the browser's time stretch, and the six rows keep the four provisional ceilings set below from 2026-11-09 and 2026-11-10. Each ceiling is the worst of its seven nights times 1.25, plus one runner hiccup, rounded up to two significant figures. The hiccup is sized from the single gaps the 2026-10-08 night put on otherwise clean rows, the fixed-250 control among them: 250 underruns/min, 700 gap ms/min, 3.3 skips/min, 280 dropped ms/min (both sized on the playhead's skip-ahead inference, and now applied to player_skips_per_min and trimmed_ms_per_min), 0.01 stalled, 0.015 silence, and one 20 ms bucket of target. Budgeted this way, every night passes when it is held out of its row's measurement except one, and that row's own ceiling includes it: the 2026-11-03 Opus bursty row raised its target to 300 ms 2.6 s in and held it for one 20 s half-life before settling at 180. Such startup holds, Opus near-zero's 100 ms on 2026-10-30 and 10-31 among them, are budgeted as measured. These ceilings record today's player as the runner hears it, not target audio quality. To re-measure, download the same artifact from the latest scheduled nightlies (`gh run download -n audio-quality-run`), which keep every run, and apply the same rule. On 2026-11-10 the inferred skip_aheads and discarded_ms keys were dropped, since they read a time stretch's catch-up as dropped audio, and every browser row gained player_skips_per_min, trimmed_ms_per_min, and av_lag_ms p50 and p95, by the same rule from the only two scheduled nights on main with today's browser player and its trim counter, 2026-11-09 and 2026-11-10 (42659882487, 42842827718). A lag is signed, so its ceiling is the worst night plus a quarter of its size plus one 20 ms bucket. These four keys are provisional until seven nights with the stretch model re-measure every row. The grade follows the model: audio rests on the shared clock, trails it through a stall, and catches up after. When a change moves a row, a steady-state lag (av_lag_ms p50) that rises, or audio the player trimmed, is a regression to fix rather than budget; a p95 lag that rises through a stall, or an underrun count that moves, is re-measured in. No ceiling is fixed at zero underruns or zero stalls; a replay row's is exactly what it measured. The replay rows are deterministic, so their ceilings are exactly what they measured, through the end of each recording, with no headroom: any regression fails them, and an improvement should be tightened in. They were last re-measured on 2026-10-22 using version 3 traces with encoded frame durations: AAC-LC uses its fixed 1024 samples, and relay-mic retains Opus packet TOC durations. relay-mic was re-recorded on 2026-11-03 from a real microphone, Chrome 156 on a MacBook Air publishing its built-in microphone through the public relay, recorded by headless Chromium on a Linux host (35 s, minimum RTT 2 ms), which raised its target ceiling from 60 to 80 ms and its silence from 0 to 0.001 as measured. The recording path itself has no added RTT. relay-mic-rtt40 is the 2026-11-03 Chrome row at 40 ms of added RTT on that laptop (180 s, minimum RTT 47 ms), kept as the baseline time stretch improves on; like every replay row, its player_underruns_per_min ceiling is what it measures, not a fixed zero. player_underruns_per_min, the player's own counter that its stats panel shows, is budgeted on the replay rows only; it counts the run-dry re-stalls the tap files as stalled time. On 2026-11-07, after the ordered read stopped delivering groups below its cursor (#8071) and took Rust's hole and age bounds (#8218), relay-bbb was re-measured at 16.1 player underruns/min, 0.025 stalled and 0.02 silence, and relay-mic-rtt40 at 8.2, its discarded audio at 70.9 ms/min and its target p95 at 320 ms rather than 780, since the stale backlog it never played no longer reaches the estimator; neither row lost audio the previous player heard. On 2026-11-08 the player began time-stretching live playback with NetEq's accelerate, deciding on the ring's depth right after each arrival and never expanding, and every replay row was re-measured with it, the browser rows left to the nightly runner. Each row's av_lag_ms, how far the playhead trails the shared clock video plays from, is what it measured: the ring rests one frame ahead of the newest arrival's playout, as the trim it replaces left it, so audio leads the clock by 19 to 23 ms at p50. Against the trim, the A/V lag p50 fell from 0.7, -17.5, -21.5, -18.2 and -15.6 ms to -22.4 on bursty-aac, -22.2 on local-aac, -22.7 on relay-bbb, -19 on relay-mic and -20.1 on relay-mic-rtt40, and its p95 fell on every row, most on relay-bbb (10.5 to -4.2) and relay-mic-rtt40 (-0.9 to -13.4). Underruns stayed where the trim had them, as the latency they would cost is not spent: player_underruns_per_min held at 4.4, 12.1, 16.1 and 6.1 and rose from 8.2 to 10.6 on relay-mic-rtt40, and relay-mic's tap now files 2 short quanta a minute, 2 ms of missing audio, where it filed none. What the stretch buys is the audio the trim dropped: discarded_ms_per_min fell from 87.8 to 0 on relay-bbb and from 127.5 to 100.6 on local-aac, whose remainder drops nothing: two receive stalls 14 and 15.3 s in leave audio 159 and 132 ms behind the clock, and the fast accelerate brings it back within half a second. player_skips_per_min and trimmed_ms_per_min, the ring's own count of re-anchors that dropped buffered audio and the media they dropped, are budgeted beside it, at 0 on every recorded row. relay-mic-stall, added 2026-11-10, is relay-mic with nothing arriving from 15 s to 16.5 s and the held frames then trickling in at twice real time until 18 s; today's player trims 2152.5 ms/min of it and its target p95 rises to 1540 ms. Once durations became whole nanoseconds, relay-bbb's delay, 200 ms plus one 1024-sample AAC frame, rounds up by a fraction of a sample, so its ring holds one more sample (23 µs) and its A/V lag p50 measures -22.6 rather than -22.7. Neither replay calibration changed the browser or native ceilings. bursty-aac is the 2026-10-18 AAC bursty isolated capture, recorded with the observer subscribed at the decoder's 2 s budget, so it holds every group the player saw. The native rows are provisional: measured on a loaded developer host (x86_64 Linux, the private PulseAudio null sink, seed 7, 60 s a row) on 2026-10-17, three runs of every row from the debug build nightly uses plus three more of the AAC rows from a release build, which played alike. Each ceiling is the worst clean run times 1.25, plus one native hiccup sized from the unimpaired Opus row's worst run (20 underruns/min, 50 gap ms/min, 2.2 skips/min, 110 discarded ms/min, 0.01 stalled, 0.005 silence, one 20 ms bucket of target), rounded up to two significant figures. On 2026-11-10, after native playback began reporting its A/V lag (#8361), the native rows' player_skips_per_min, trimmed_ms_per_min, and av_lag_ms p50 and p95, and every key of the native and browser congestion rows, were measured by the same rules from three runs of each on a loaded developer host (x86_64 Linux, the debug build, seed 7, 60 s a row), with no void; native audio rests 9 to 95 ms behind its clock at p50. These are provisional until the stretch model's nightly re-measure, which also moves the browser congestion rows onto the runner. On the 2026-10-17 host the null sink's clock ran about 150 ppm slow against the monotonic clock, read within 5 ms at p95 per sample, and its output latency sat between 14 and 28 ms. Five of the 36 AAC rows (none of the 18 Opus rows) voided on the clock: the device stopped pulling for about two seconds mid-row, sometimes after a backend error and a stream restart, which is a native playback defect rather than noise, and the ceilings exclude those runs. Re-measure from the nightly runner's artifacts once it has a week of native rows. On 2026-11-11 every player began following one stretch rule on the shared clock: it slows down below its target, speeds up above it, and skips only past its bound, which in the browser is the subscription's max delay. Every replay row was re-measured with it, the browser and native rows left to the nightly runner. Audio now rests on the clock rather than one frame ahead of it, so the A/V lag p50 rose from about -20 ms to 0.4 ms on relay-bbb, 1.1 on relay-mic, 1.2 on relay-mic-rtt40 and 1.8 on local-aac. Its p95 rose on every row, since audio that falls behind is now caught up by speeding up rather than trimmed: the shared clock steps forward as its reference converges through a tune-in backlog, and again when a target falls, and audio follows at the accelerate's pace, which on the replay's constant tone is one 2.5 ms period a window, slower than on speech. bursty-aac rests 5 ms behind its clock, so its p50 of 10.4 ms is mostly its catch-ups, and relay-mic-stall's p50 of 66.1 ms is the catch-up after its target falls from 1540 to 80 ms, which outlasts half the row. In exchange nothing is trimmed: relay-mic-stall's trimmed_ms_per_min fell from 2152.5 to 0, its skips from 6.1 to 0, and its underruns from 24.2 to 0 a minute, and player_underruns_per_min fell from 12.1 to 4 on local-aac, 16.1 to 0 on relay-bbb, 6.1 to 0 on relay-mic, 10.6 to 2.7 on relay-mic-rtt40 and 6.1 to 4 on relay-mic-stall, holding at 4.4 on bursty-aac. The tap now files 3.3 to 6.1 short quanta a minute on the AAC rows, under 0.3 ms of missing audio, where it filed none. To deliver the audio a stall delays, the container consumer now waits and skips at the subscription's max delay, as native already does, rather than at the delay plus lookahead. relay-mic-rtt40's join backlog arrives reordered, so the consumer now waits on its holes and the estimator observes the frames that fill them, up to 775 ms late, and its target p95 returns from 320 to 780 ms, held for one half-life after the join. Later on 2026-11-11 audio began skipping as far as the shared clock's reference moves earlier, once those moves add up to one stretch window and only as far as it holds audio past its target, rather than speeding through them, and every replay row was re-measured. The tune-in backlog's jumps now land before the graded window, so the A/V lag p95 fell from 453.1 to 2.1 ms on relay-bbb (its p50 from 0.4 to -1.3), 743.7 to 1.9 on relay-mic-rtt40, whose p50 fell from 1.2 to 0.4 and its tap's missing audio rose from 0.6 to 0.9 ms a minute, since a jump past what is decoded skips the rest once it lands,, 335.5 to 39.2 on local-aac and 88.6 to 0.3 on relay-mic, and the p50 from 66.1 to 25.3 on relay-mic-stall, with nothing trimmed in any graded window. The tap's underruns fell from 6.1 to 2 a minute on local-aac and 6.1 to 0 on relay-bbb; local-aac's silence rose from 0.001 to 0.002. Later on 2026-11-11 the container consumer began waiting on a missing group only for the delay plus the lookahead, while still delivering groups that arrived until they age past the max delay, and every replay row was re-measured. relay-mic-rtt40's target p95 fell from 780 back to 320 ms, since the holes in its reordered join backlog no longer delay the frames behind them; its tap's missing audio rose from 0.9 to 1.1 ms a minute and its A/V lag p50 and p95 from 0.4 and 1.9 to 0.7 and 2.4 ms. bursty-aac's silence rose from 0.007 to 0.009 and its A/V lag p95 fell from 162.1 to 156.3 ms, and relay-bbb's lag p50 and p95 moved from -1.3 and 2.1 to -1.2 and 2.2 ms. relay-mic-stall did not move. Also on 2026-11-11 the stretch became a player setting, 2 s by default, and the audio subscription's max delay became the delay plus it rather than the 2 s constant, so the container consumer delivers a late group that much longer before it ages out, and every replay row was re-measured: only relay-mic moved, its A/V lag p50 and p95 from 0.3 to 1.6 ms, under the 2.5 ms pitch period the stretch can move it by. Holding the constant max delay with the new bound reproduces the old values exactly. On 2026-11-12 the replays stopped playing a constant and began playing a seeded speech-like signal (syllables gliding a 90 to 220 Hz fundamental, with pauses held above the silence classifier), since the stretch's pitch search found only its shortest period in a constant, and the browser ring stopped zero-filling a hole of up to a millisecond and two samples between frames, as native stopped writing silence for one: the recorded AAC timestamps land up to 2.5 samples either side of a 1024-sample boundary, and on speech the 2-sample holes this filled ended 119.1 quanta a minute short on bursty-aac. Every replay row was re-measured. The tap's underruns fell to 0 on bursty-aac (from 3.3) and local-aac (from 2), and relay-mic-rtt40's missing audio from 1.1 to 0.3 ms a minute while its player underruns rose from 2.7 to 3.1. The A/V lag p95 fell from 156.3 to 104.8 ms on bursty-aac, 39.2 to 15.1 on local-aac, 2.2 to 1.4 on relay-bbb and 501.8 to 484.9 on relay-mic-stall, whose p50 fell from 25.3 to 1.1, and rose from 1.6 to 3.4 on relay-mic and 2.4 to 3.5 on relay-mic-rtt40. Where audio rests moved within one speech pitch period, which is as close as one stretch seam can steer it: the p50 is 11 ms on bursty-aac (from 10.4), -4.2 on relay-bbb (from -1.2), 1.1 on relay-mic (from 1.6), 0.9 on relay-mic-rtt40 (from 0.7) and 1.7 on local-aac (from 1.8). On 2026-11-13 the stretch default fell from 2 s to 200 ms and audio trailing the clock past it began skipping onto the clock rather than back to the bound, as video skips, with the audio subscription still reaching 2 s past the delay; every replay row was re-measured. A target fall past 200 ms now costs one skip rather than seconds of catch-up, and so does a trail past it, and a rise past it pads silence until the clock arrives rather than slowing down: relay-mic-stall skips the 1.46 s it trails after the outage and its 1540 to 1020 and 1020 to 80 ms falls, and pads the 0.5 s its 60 to 1540 ms rise leaves it ahead, where it slowed down for 4 s, so it trims 4678.2 ms a minute where it trimmed none and its stalled share rose from 0.05 to 0.067 and silence from 0.049 to 0.085. The frames it holds through the outage then trickle in at twice real time, so its low buffer slows it to 154 ms behind the clock, where it had sat ahead of it: its A/V lag p95 fell from 484.9 to 118.6 ms and its p50 rose from 1.1 to 2.2. relay-mic-rtt40 skips its 320 to 80 ms fall, 0.3 skips and 83.3 trimmed ms a minute, its p95 from 3.5 to 2.9 ms. bursty-aac skips a 225 ms trail 6 s in, 1.1 skips and 312.7 trimmed ms a minute, its p95 from 102.7 to 49.2 ms and p50 from 11 to 10. relay-mic trims nothing, but its tune-in's skips land on the clock rather than at the bound, which moves where it rests within a pitch period: its p50 from 1.1 to 2.2 ms and p95 from 3.4 to 4.5. Trimming at a target fall or a trail past 200 ms is the rule, not a regression. Also on 2026-11-13, for the same change, the congestion rows' trimmed_ms_per_min ceilings were raised by the same rule from two runs of each on a loaded developer host (x86_64 Linux, the debug build, seed 7, 60 s a row): to 3800 on chromium AAC (worst 2740.3 ms/min), 3700 on chromium Opus (2697.2), 3900 on native AAC (2996.9) and 4600 on native Opus (3532.3), whose player_skips_per_min rose to 15 (9.8). The trim is the outage's audio skipped onto the clock rather than played seconds late: against main on the same host, chromium's A/V lag p95 fell from 1396.8 to 61.6 ms on AAC and 1424.8 to 64.3 on Opus, and native's from 1246.3 to 116.7 and 826.9 to 117.5. These four ceilings are provisional until the nightly re-measures the congestion rows. Also on 2026-11-13, the estimator began counting one outage, the intervals its stalled frames span and the empty ones before them, as one observation, and every replay row was re-measured on top of the snap: only relay-mic-stall moved. Its single stall no longer raises the target, so its target p95 fell from 1540 to 80 ms and the trail the outage leaves is skipped onto the clock at once rather than through the target's falls: 2 skips a minute where it took 6.1, 5122.3 trimmed ms a minute where it trimmed 4678.2, and its A/V lag p95 from 118.6 to 19.7 ms. With 80 ms of buffer rather than 1540, more of the outage plays as silence: its stalled share rose from 0.067 to 0.092 and silence from 0.085 to 0.092. On 2026-11-14 every player began following the clock by its lag alone: a continuous WSOLA time stretch for any content, at a rate proportional to the lag and capped at 10%, which leaves a lag within half a 10 ms hop alone, and slows down while the audio received ahead of the playhead is about to run out, replacing NetEq's buffer-level decision and pitch-period splices. Every replay row was re-measured. Audio now rests within half a hop of the clock rather than within a pitch period: the A/V lag p50 fell from 10 to 0.2 ms on bursty-aac, whose p95 fell from 49.2 to 0.2, from 1.7 to -1.3 on local-aac, 2.2 to 1.8 on relay-mic and 0.9 to 0.6 on relay-mic-rtt40, and moved from -2 to 2.7 on relay-bbb and 2.2 to 3.1 on relay-mic-stall, each within the half hop the rule leaves alone. The p95 rose from 3.6 to 6.2 ms on relay-bbb, 4.5 to 6.9 on relay-mic, 2.9 to 6.9 on relay-mic-rtt40 and 19.7 to 32.1 on relay-mic-stall, and fell from 15.1 to 14.3 on local-aac: a lag under a hop is corrected only once its pace adds up to one. player_underruns_per_min rose from 4.4 to 6.6 on bursty-aac and 4 to 8.1 on local-aac, whose recorded outages outlast what slowing down at 10% bridges, and fell from 3.1 to 1 on relay-mic-rtt40, whose tap's underruns fell from 0.7 to 0.3 a minute; bursty-aac's silence rose from 0.007 to 0.008. Nothing more is trimmed on any row. The browser and native rows were measured on a loaded developer host (x86_64 Linux, the debug build, seed 7, 60 s a row) against main the same day and left to the nightly runner: every browser row's A/V lag p50 now sits between 27 and 39 ms, where the fixed-250 control does on that host, where wide and bursty sat up to 58 ms behind, and every native row's within 15 ms of the clock. Also on 2026-11-14 the estimator stopped folding the intervals a subscription's join spans, those holding a frame due before its backlog finished arriving, and every replay row was re-measured: only relay-mic and relay-mic-stall moved. Their first frame arrives slower than the path after it, so their first interval folds nothing and the target leaves the prior 500 ms later, which moves where audio rests within the half hop the rule leaves alone: relay-mic's A/V lag p50 from 1.8 to -0.3 ms and p95 from 6.9 to 9, and relay-mic-stall's p50 from 3.1 to 1.7. relay-mic-rtt40 did not move: the browser consumer closes its join's stale groups below the cursor unread, so the estimator never sees when that backlog ended, and its target p95 held at 320 ms. Later on 2026-11-14 both container consumers began reporting when each group they close unread arrived, if it is at or below the publisher's largest group as the subscription began, which extends a join still open. On 2026-11-15 the traces gained that largest group (version 4), so the replay's consumer learns it as a wire subscriber does, and relay-mic-rtt40 was re-measured: its target p95 fell from 320 to 120 ms, underruns per minute rose from 0.3 to 0.7, player skips per minute fell from 0.3 to 0, trimmed ms per minute from 83 to 0, A/V lag p50 rose from 0.6 to 2.3 ms and p95 from 6.9 to 8.6, the trade of reporting the join's backlog. No other replay row moved. On 2026-11-17 the browser player began reading audio through the ordered reader, which waits for a hole as long as the ring could still play it and reports each half-open range of keys it gives up on as one gap, and the replay dropped the trace's largest group: a second subscription now observes every arrival for the estimate, the join's backlog included. Every replay row was re-measured. bursty-aac's player underruns per minute fell from 6.6 to 4.4 and silence share from 0.008 to 0.007; audio rests at 0.1 ms until a stall at 19.5 s, from which the stretch catches up to 4.2 ms behind the clock, inside the half hop it leaves alone, so its A/V lag p50 and p95 rose from 0.2 to 4.2 ms. relay-bbb's A/V lag p50 fell from 2.7 to 1.4 ms and p95 rose from 6.2 to 7.6. relay-mic-rtt40's underruns per minute fell from 0.7 to 0.3, as did its player underruns per minute, underrun ms per minute from 0.8 to 0.2, A/V lag p50 from 2.3 to 0.6 ms, and p95 rose from 8.6 to 9.7, its target p95 holding at 120 ms; 0.6 ms of that rise is the replay's ordered reader taking the catalog's warmup into its budget, as the browser's does. No other replay row moved. On 2026-11-18 every browser row's av_lag_ms_p50 ceiling became 5 ms, the half hop the stretch leaves alone, rather than a measurement: where a row rests inside it is chance, which the four scheduled nights since the stretch reached the runner (43724172821, 43909101837, 44127032469, 44342367698, 2026-11-15 to 11-18) put between -3 and 4.3 ms on AAC. The Opus rows rested 2.7 to 11.1 ms behind on those nights with no acceleration for a minute at a time, since the browser mapped the clock onto its ring a 6.5 ms pre-skip late, so Opus really played that far behind video; with the ring played at presentation time, as the hang draft and the native player have it, they rest inside the same window. A browser row resting past 5 ms is drifting audio the stretch failed to correct, so it fails rather than being budgeted in.", "rows": [ { "runtime": "chromium", @@ -13,7 +13,7 @@ "stalled_share": 0.017, "silence_share": 0.023, "target_ms_p95": 450, - "av_lag_ms_p50": 190, + "av_lag_ms_p50": 5, "av_lag_ms_p95": 200 }, { @@ -28,7 +28,7 @@ "stalled_share": 0.01, "silence_share": 0.015, "target_ms_p95": 340, - "av_lag_ms_p50": 8, + "av_lag_ms_p50": 5, "av_lag_ms_p95": 15 }, { @@ -43,7 +43,7 @@ "stalled_share": 0.01, "silence_share": 0.015, "target_ms_p95": 290, - "av_lag_ms_p50": 11, + "av_lag_ms_p50": 5, "av_lag_ms_p95": 17 }, { @@ -58,7 +58,7 @@ "stalled_share": 0.01, "silence_share": 0.015, "target_ms_p95": 290, - "av_lag_ms_p50": 7.5, + "av_lag_ms_p50": 5, "av_lag_ms_p95": 14 }, { @@ -73,7 +73,7 @@ "stalled_share": 0.022, "silence_share": 0.023, "target_ms_p95": 300, - "av_lag_ms_p50": 61, + "av_lag_ms_p50": 5, "av_lag_ms_p95": 82 }, { @@ -88,7 +88,7 @@ "stalled_share": 0.033, "silence_share": 0.042, "target_ms_p95": 400, - "av_lag_ms_p50": 25, + "av_lag_ms_p50": 5, "av_lag_ms_p95": 85 }, { @@ -103,7 +103,7 @@ "stalled_share": 0.01, "silence_share": 0.015, "target_ms_p95": 400, - "av_lag_ms_p50": 17, + "av_lag_ms_p50": 5, "av_lag_ms_p95": 32 }, { @@ -118,7 +118,7 @@ "stalled_share": 0.01, "silence_share": 0.015, "target_ms_p95": 340, - "av_lag_ms_p50": 15, + "av_lag_ms_p50": 5, "av_lag_ms_p95": 21 }, { @@ -133,7 +133,7 @@ "stalled_share": 0.01, "silence_share": 0.017, "target_ms_p95": 200, - "av_lag_ms_p50": 15, + "av_lag_ms_p50": 5, "av_lag_ms_p95": 23 }, { @@ -148,7 +148,7 @@ "stalled_share": 0.01, "silence_share": 0.015, "target_ms_p95": 150, - "av_lag_ms_p50": 15, + "av_lag_ms_p50": 5, "av_lag_ms_p95": 21 }, { @@ -163,7 +163,7 @@ "stalled_share": 0.022, "silence_share": 0.019, "target_ms_p95": 200, - "av_lag_ms_p50": 70, + "av_lag_ms_p50": 5, "av_lag_ms_p95": 93 }, { @@ -178,7 +178,7 @@ "stalled_share": 0.01, "silence_share": 0.018, "target_ms_p95": 620, - "av_lag_ms_p50": 28, + "av_lag_ms_p50": 5, "av_lag_ms_p95": 70 }, { @@ -193,7 +193,7 @@ "stalled_share": 0.11, "silence_share": 0.11, "target_ms_p95": 850, - "av_lag_ms_p50": 1.3, + "av_lag_ms_p50": 5, "av_lag_ms_p95": 44 }, { @@ -208,7 +208,7 @@ "stalled_share": 0.078, "silence_share": 0.082, "target_ms_p95": 1300, - "av_lag_ms_p50": 8.3, + "av_lag_ms_p50": 5, "av_lag_ms_p95": 32 }, {