[fishloa]

Closes #91. Delegated to deepseek-v4-flash via crush; $0.04 for the session. I verified every claim independently — details below.

Root cause — two compounding faults, both measured

1. The feed burst overflowed the worklet's 16 s ring. _pumpAudioInner drained all decoded PCM and posted it to the worklet in one tick. A clip decodes far faster than it plays, so the ring overflowed and silently dropped audio. Measured: audioSec froze at 16.00 with ahead=2.87 — about the ring's capacity, i.e. 2.87 s of fed audio physically unplayable.
2. done was never reached while buffered audio was still queued. The end-of-stream path flushed and emitted ended while roughly a segment remained posted-but-unplayed. And done was a transient on the emitted stat object, so a later stats emission cleared it — the first observed run froze at audioSec=14.88, done=false, permanently.

That second half is the worse one: a host could not distinguish "stream finished" from "player froze".

Fix

• Pace posting — PCM is drained into a JS-side queue and reposted incrementally, capped at _AUDIO_LEAD_S of play-ahead, so the ring cannot overflow.
• Drain to completion — a new end-of-stream drain keeps posting queued PCM and waits for play-ahead to reach ~0 before reporting the end.
• Persistent done — set as a field on _stats so a later emission cannot clear it; ended fires exactly once.

Verification — I re-ran all of it, and negative-controlled the tests

gate	result
cargo fmt --all --check	clean
clippy --workspace --all-targets -- -D warnings	zero
cargo nextest run --workspace	157/157
bun test (packages/player)	60/60
tsc --noEmit -p packages/player	exit 0
bun run test:streams	21/21 (14 existing + 7 new)

The new tests genuinely bite. I reverted only skyfire-player.js to its pre-fix state, kept the tests, and re-ran:

unfixed:  7 failed   (1.8–2.1s each)
fixed:    7 passed   (19–20s each)

The timing difference is the tell: unfixed, the tests short-circuit in ~2 s because the old code declared the stream over almost immediately. Fixed, each takes the clip's full ~19 s because it now actually waits for playout. A test that passed by accident could not show that asymmetry.

Measured on arte (advertised 18.76 s): before 14.88–16.01 s played with done=false; after 18.85 s, done=true, ended once, ahead→0.03, queued=0. Within 0.1 s of advertised, against a 1 s criterion.

Scope

Additive only — packages/player/skyfire-player.js and 70 new lines in web/tests/streams.spec.mjs. No existing test was modified, which I checked by diffing rather than trusting the report.

Note on the oracle

The assertion is against the playlist's own advertised total (sum of EXTINF), computed per-stream at test time, not a hardcoded number. That was chosen deliberately so it cannot be satisfied by tuning a constant, and it covers all seven fixtures rather than the two I originally measured.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Qsz3AufTr5z6MPZJzkoSCT