[ingoau]

Part of #228 (desktop only for now: hocket-android builds hocket-core without the decode feature, so the scanner measures nothing on Android).

compute_replaygain (a complete EBU R128 scan) existed but nothing called it: downloads only copied the server's track.replay_gain, so tracks without ReplayGain tags — and transcodes that drop them — played with no level correction, which also made AutoMix jump in level mid-blend.

What this does

• New GainScanner trait in the downloads module; the real implementation calls compute_replaygain (feature decode), injectable for tests.
• Background worker: Downloads::new spawns one thread that takes (server, track) requests from a channel, deduplicated while pending, and exits when the Downloads is dropped. Nothing is measured on the playback or pin path.
• Two scheduling points: a pin download completes without a server track gain (fetch_track), and a stream-cache entry is promoted to complete (cache_merge).
• Startup catch-up: the scan queue lives in memory, so files that completed without being scanned (process died mid-scan) are found again on the next start; the worker's own skip checks make this cheap.
• run_gain_scan skips tracks the server tags or that already have a stored gain, measures the pin's file (else the newest complete cache entry), and stores the result via set_gain (pin rows) or the new cache_entries.gain_db/peak columns (migration 0010). A decode failure stores nothing; the startup catch-up retries it on the next launch.
• Downloads::gain_with_peak (and gain on top of it) reads a download's value, else the newest measured cache entry (original preferred over a transcode), so resolve/media_source_for level untagged tracks that were never pinned. In media_source_for, the measured track gain fills what the server left out — a partial tag set (album gain, a bare peak) no longer shadows the on-device scan, and the measured peak reaches the clip clamp.
• Adoption keeps its gain: a cache file adopted as a pin's download carries its measured gain onto the pin rows instead of being re-measured.
• The module doc and the compute_replaygain doc now match reality.

Tests

Five new tests around an injected fake scanner: untagged download measured once (with peak, and resolve picks the gain up); a server-reported gain never triggers a scan; a complete cache entry is measured for a non-pinned track (stored on the cache row, resolve picks it up); an undecodable file stays gainless without rescans; and a fresh process scans files completed while it was away.

cargo test -p hocket-core: 877 passed. Verified with and without the decode feature, fmt and clippy clean (only pre-existing warnings remain).

Generated by GLM-5.3-Flash in OpenCode