feat(audio): warm-up gate on the adaptive NMI-rate loop to kill the start/seek pitch glide

[kfox]

Problem

On the host-DMA audio path, the adaptive NMI-rate loop seeds its R-rate EMA off the first measurements after the consumer starts — but those are unrepresentative. Right after a start/seek the video pipeline's bus load hasn't reached steady state yet (post-seek decode catch-up plus the playlist's frame-drop snap-forward), so the measured consumer rate R reads high (~11.5k Hz) until the steady-state VIC/DMA tick-loss arrives a few seconds later and R settles (~10.1k Hz).

The loop chased that transient — stepping the CIA #2 latch away from its seed and then all the way back — producing an audible start-of-playback pitch glide. The motion is pure waste: for a bitmap scene the seed latch is the ceiling, which is already the converged value, so holding the seed steady would have landed dead-on.

Fix

Add a warm-up gate to the rate loop. For NMI_RATE_LOOP_WARMUP_S (3.0 s) after the consumer starts — and after a large playback disturbance — the loop keeps feeding R into its EMA but holds the latch at the near-converged seed, then makes its first decision from a settled estimate. The seed plays at ~the correct rate during the hold, so there's no glide either way.

Arming points:

• Consumer start — armed in the worker's post-prebuffer reset (covers the initial seek).
• Large frame-drop — the playlist calls the new AudioStreamer.note_playback_disturbance() when a deadline snap-forward drops more than _AUDIO_DISTURBANCE_DROP_S (0.5 s) of frames (a seek catch-up or a stream rebuffer). Routine 1–3 frame drops stay below the bar.
• Mid-stream display-mode change — the bus-halt profile (hence R) shifts with the mode, so the new seed is held until the new load settles.

The EMA is left intact across a disturbance (not re-seeded) so it keeps tracking through the gap. Steady-state behavior is unchanged — only the start/seek/disturbance transient is affected. Host-DMA path only (REU pump has its own C64-side governor; open-loop/static paths are unaffected).

Tests

• Audio: warm-up holds a would-be latch step, still warms the EMA during the hold, releases and steps after the deadline, and is re-armed by note_playback_disturbance().
• Playlist: a large drop signals the audio loop, a small drop does not, and the audio=None case is safe.

make lint typecheck test all clean (ruff, mypy --strict, pyright, full unittest suite).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 79.56%. Comparing base (77067af) to head (8ea8e27).
⚠️ Report is 2 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #39      +/-   ##
==========================================
+ Coverage   79.55%   79.56%   +0.01%     
==========================================
  Files          68       68              
  Lines       12847    12858      +11     
  Branches     1896     1898       +2     
==========================================
+ Hits        10220    10231      +11     
  Misses       2188     2188              
  Partials      439      439              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.