[AlxcBTW]

Two flaky tests that fail on the macOS runner and turn unrelated PRs red (#110-#112).

Causes

the_streamer_s_sound_is_played_by_the_viewer (sound.rs): asserted that the single highest played sample is below 0.35 (tone peak 0.25). Opus overshoots on the first packets and where a late packet is concealed or rebuilt, and a slow shared runner stalls the pipeline often enough to hit that (seen: peak 0.388). The test now checks what it is meant to prove, the level of the tone: the 90th percentile of the audible samples (expected 0.247) must lie in 0.2..0.28, and no sample may reach 0.5 (about double gain). Not a retry, and a wrong gain, silence or garbage still fails.

a_stalled_receiver_costs_a_bounded_queue_not_a_task_per_packet (traffic.rs): expected at least one expired packet after 10 000 sends and 1.5 s. The sender only stalls once the connection's stream limit/windows are used up, and how many of the 10 000 packets reach the queue depends on how the sender task and the send loop interleave. On a slow runner the sender keeps up with the burst, never stalls, nothing expires (log: tasks 15/15/15, expired == 0). The test now keeps feeding paced batches until a queued packet expires (up to 20 s), checks the task bound all the while, then waits the 1.5 s as before. The thing proven is unchanged: bounded queue, no task per packet, expiry on a stalled receiver.

Failure rates before (last 100 runs of ci, job check (macos-latest), 50 runs)

• 20 of 50 failed (40 %).
• the_streamer_s_sound_is_played_by_the_viewer: 8 (16 %).
• a_stalled_receiver_... : seen in run 40495192564 (not in the window counted).
• Other macOS failures, not touched here: real_window::the_worker_saves_a_clip_of_a_real_window (share-worker access.rs) 3, share_play.rs:103 (a_game_that_goes_away_ends_the_share, hand_over_...) 2, tunnel.rs:451 1, clippy errors on unmerged heads 3 (and 1 other not compiled).

Stability

macOS job rerun results are listed in a comment below.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[AlxcBTW]

Stability: the macOS jobs (check and devices) of the final head were rerun 3 more times after the first green run: 4 of 4 green for both. (Devices once failed before on tunnel.rs:451 and play.rs, both unrelated and green on rerun.)