Root causes found and fixed:
- QueueManager could panic and kill the playback task permanently:
is_last_track() underflowed on an empty queue, remove_tracks accepted
pos == len (Vec::remove panic) and corrupted positions when removing
multiple tracks (indices shifted mid-loop), shuffle_behind indexed
out of range on an empty play order, insert_tracks shifted play-order
entries by the queue length instead of the inserted count and then
assert!()ed on the resulting inconsistency, and clear() left
play_order stale. All mutation methods are now guarded, multi-remove
works highest-position-first, and an inconsistent play order is
rebuilt instead of panicking. Regression tests cover these cases.
- The tidal access token was only obtained at startup and never
refreshed, so long-running sessions ended with every track fetch
failing (playback just stopped at the next track boundary). Login
state now lives behind a lock; tokens are refreshed proactively
before expiry (5 min margin) and once reactively on a 401, and all
API responses are status-checked (new ClientError::ApiError) instead
of being fed to the JSON decoder blind. The http client also got a
30s timeout so a hung connection cannot wedge the provider loop.
- (from the rodio rewrite, same bug class) end of stream used to be
detected by string-comparing an io::Error message; any other decode
or network error ended the stream silently without an EndOfStream
message, so playback never advanced. EOS is now a guaranteed
callback with a generation counter.
Plus workspace-wide clippy cleanup (zero warnings), cargo-machete
cleanup, and fmt.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
d504ebc85f