The conversion thread relied on `push_chunk` returning `Err` to notice the
track had changed — the comment said so in as many words ("Reader gone (track
changed/stopped) — just stop"). That check could never fire: `push_chunk`
only extends a Vec and fails solely on a poisoned mutex, and `BufferState`
carries no reader-gone flag and no Drop.
So skipping a DSD track left the thread converting to the END of the file,
growing a buffer nothing would ever read. A 60-minute DSD64 album track is
~1.9 GB of 88.2 kHz/24-bit PCM held for nothing, on a box that hard-freezes
under memory pressure.
The fix reuses the contract that already exists for the CMAF feeder:
`register_stream_feeder` with the play generation, plus a cooperative check
once per converted block. Registration happens AFTER the spawn and carries
the generation the play was born under, so a supersede that lands in between
cancels the thread instead of registering it. Like the network feeder's own
cancel, it neither errors nor completes the buffer — the outgoing engine keeps
playing what is already buffered until the new source is installed.
Confirmed in a live session log, not inferred:
[CMAF-STREAM] Cancelling feeder for superseded track 5672
Player: DSD conversion for superseded track 5672 cancelled
cc