The visualizer's PCM cache decoded the entire episode into RAM (22050 Hz
mono s16 ~160 MB/hr of audio) and held it until stop() — a 3-hour episode
pinned ~500 MB and long-form content hit 2.5 GB. The 4x decode also pulled
the whole remote file even when only minutes were listened to.
- audio-pcm-cache: sliding window around the playback position — the
decode head caps at maxAheadSec (600s) ahead of the cursor, segments
older than keepBehindSec (300s) are pruned, and the tail refills as
playback advances. Steady state ~40 MB regardless of episode length;
a backward seek past the window restarts a segment there (the existing
seek-hole mechanism, no new failure mode).
- feed: cap the full-parse episode cache at 1000 episodes/feed so
archive-heavy subscriptions can't pin their entire history in RAM;
the visible list stays bounded by the user's cache preference and
fetch-more keeps working within the ceiling.
- tests: pin the new head-cap and prune contracts (8/8 in
audio-pcm-cache.test.ts; full suite 193 pass).
Also includes the in-flight cleanup/refactor pass (cover-art resolve
helper, page and comment tightening, ESLint config removal).
ensureDecodeAround let a running decode pass close ANY forward gap in
place — at 4x pacing, skipping 30 min ahead meant ~7.5 min for the
frontier to arrive: bars held their last frame indefinitely. Now a gap
beyond 15s restarts the pass at the seek target (network coverage there
in ~1.6s, verified); smaller gaps close in place (cheaper than a
reconnect + range request). Seek-key holds are debounced (400ms) so
rapid re-seeks don't reconnect-spam the stream's server, and pending
seek-decode cancels on pause/stop so no ffmpeg restarts while paused.
Also fixes a pre-existing config-write race that intermittently failed
visualizer-toggle.test.ts: updateConfig re-resolved the config path and
re-read the patch state when the deferred write-chain drained, so a
queued save could land in a directory XDG_CONFIG_HOME had since been
pointed at (or carry state mutated after queueing). Path and patch
snapshot are now captured eagerly at call time.
Two fragility points, rebuilt at the root:
Playback: one resident mpv daemon (--idle --keep-open) with a persistent
IPC connection and observe_property state instead of spawn-per-episode and
connect-per-poll. Play/pause/seek are sub-ms commands; time-pos pushes at
~20Hz; external pauses arrive as events. Boot session restore preloads the
episode paused (loadfile + paused time-pos seek, since mpv defers --start
stream work until playback) so first Play is a ~400ms unpause instead of a
cold 4.3s open+seek. Load ops are mutex-serialized so a raced preload
cannot clobber an in-flight play.
Data throttling: mpv demuxer cache capped (cache-secs=90, max-bytes=40MiB)
so a paused preload no longer races to its 150MiB default (measured
45.7MB/12s); decoder paced at 4x realtime instead of 84x so playback start
isn't starved by the visualizer ripping the whole episode.
Visualization: replaced the paced-ring reader (AudioStreamReader) with a
position-indexed PCM cache (audio-pcm-cache). ffmpeg fills a cache indexed
by absolute playback time; reads at the player position are always exact.
Pause freezes the render loop, resume re-arms it — no coverage guessing,
no clamped-buffer freeze (the pause->broken-waveform->freeze bug). Seeks
and speed changes need no pipeline restarts; uncovered reads return empty
and the last frame holds.
Cover art: persistent per-URL disk cache under XDG cache dir; play() no
longer awaits a curl subprocess (up to 8s). Cache hit = one stat; misses
apply late via mpv video-add.
Test suite: 161 pass. New tests pin the position-index contract (sample-
exact window reads, hold-on-uncovered, pause-keeps-cache, seek segments),
the daemon contract (play/pause/resume/seek/stop, preload fast path,
EOF->replay), and cover cache/single-flight/404.
