[ingoau]

Fixes #214.

Cause

When the core stopped, the downloader task was cancelled: reader handles dropped, fetch returned cancelled, and the task was dropped mid-backoff or by runtime shutdown. Nothing set closed, and cancelled went through record_fetch_error as if transient — so the decoder blocked in HttpSource::read logged "waiting through fetch error" every 200 ms forever, and shutdown blocked on the decode thread.

Fix (crates/hocket-core/src/audio/native/http.rs)

• CloseGuard: a drop guard in downloader sets closed and notifies data_ready whenever the task exits, however it exits — error return, task cancellation (engine stop), or runtime shutdown. Any decoder blocked in read wakes and surfaces an error.
• cancelled is final: is_final_fetch_error treats FetchError::Cancelled as non-retryable. The downloader returns instead of backing off and retrying, and read surfaces the error ("cancelled") instead of waiting through it. A cancelled body-stream error stays retryable (the stream reader also cancels handles it evicts as idle after a long pause); the refetch then fails as cancelled after a real stop and ends the downloader.

Tests

• cancelled_fetch_surfaces_from_read_instead_of_waiting: body cancelled after the first chunk → read returns an error containing "cancelled" promptly.
• runtime_shutdown_wakes_a_blocked_decoder: fetch stalls forever, decoder blocked in read_to_end on another thread; dropping the runtime (engine stop) wakes it and the thread exits.
• Both verified to fail/hang against the old behavior.

Generated by GLM-5.3-Flash in OpenCode