android: make playback and prefetch recover instead of stall

[ingoau]

Why

The Android app sometimes stopped playing without saying why, and sometimes did not cache the upcoming tracks. A pass over the ExoPlayer bridge, the service lifecycle and the core's playback and prefetch handling found several ways to end up silently stuck or to give up on caching.

What changed

ExoBackend (Android)

• Paused is reported only when the player no longer means to play (playWhenReady false). A stall with it still set (a rebuffer, a retried load, an error being recovered from) is Buffering. Before, such stalls showed the track as paused and let the service count the player as idle and stop itself mid-recovery.
• Errors a later prepare() can fix are widened from connection failures/timeouts to 5xx statuses, a stuck player and core stream handles that went away under the player (the core closes idle handles after ten minutes, and ExoPlayer legitimately idles for minutes once its buffer is full).
• While offline the retry waits instead of spending attempts; the network coming back (NetworkMonitor -> ExoBackend.onConnectivityChanged) retries at once with a fresh schedule.
• CoreStreamLoadErrorPolicy retries a closed/reaped handle with a fresh open at the same position, which the cache serves.
• The service's idle stop and task-removed stop never release a player that is still buffering or recovering (ExoBackend.isBusy).

Core

• An item adopted at a gapless boundary that the backend never confirms with TransitionedToNext (a SetNext still in flight when the player ran out, a preload it dropped) is loaded explicitly after two seconds instead of leaving the session on an item nothing plays.
• A preloaded follow-up that fails before the transition is forgotten rather than restarting the current item from the top.
• Prefetch retries failed fetches with a growing backoff (15 s to 5 min), immediately on a network change, and treats a read that left no cache entry as a failure instead of starting it over at once.
• The stream reader drops reqwest's read timeout, which also ran while ExoPlayer's full buffer left the body unread for minutes (a spurious network error once per load burst on long tracks, and a re-transcode for transcoded streams). It now waits for the server only while a read is actually pending.

Tests

• Core: cargo test -p hocket-core --features sim (three new actor tests: adoption fallback, prefetch backoff retry, prefetch retry on network return); clippy and fmt clean.
• Android: :core and :playback unit tests (four new: error classification, stall reporting, busy state, closed-handle retry policy); the real-core end-to-end JVM test against the host build of the core.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DnhqP7bbeXKtuNP9dkcrPy

---

Generated by Claude Code