[michaelm-afk]

Audio Sync could resume playable PCM through a second producer fade, or stay unprimed merely because its queue was below target. After actual depletion it could resume on a tiny block instead of rebuilding the target reserve. These discontinuities caused audible clicks in the tested Tomba EU1 route.

This follows merged #769 without rewriting it:

• Preserve playable queued PCM and its read cursor on sink resume; re-prime only after actual depletion.
• Join newly resumed source PCM over 3 ms and avoid a second producer fade at the retained-PCM seam.
• Count depletion episodes once and rebuild the existing target after starvation. Normal steady playback, guest clocks, SPU/HQ mixing and presentation pacing are unchanged.
• Extend the existing Audio Sync API with a validated 20–180 ms / 5 ms next-session target. Enabled default is 70 ms; disabled keeps the configured fallback. Titles read committed mod options before device creation; no live retargeting or adaptive buffer policy is introduced.
• Optional producer-thread diagnostics report configured/active target, fill, depletion/overflow and host device lifecycle counters at most once per second. No callback file I/O or title-specific logging paths.

The activation plugin, saved mod options and default-on choice belong to the title, not the engine. Consumer: mstan/TombaRecomp#55. Paired UI: RetroPortingToolKit/recomp-ui#122; it uses the existing ModProvider storage. This API/continuity correction can be reviewed independently of the UI.

Validation

On current master ffbd9d1, native Windows Clang Release:

• Full psx-runtime build with the paired current-master recomp-ui succeeds (OpenGL; netplay, rewind, setup wizard and debug tools disabled; Vulkan SDK absent/stub).
• audio_sync_target_test, audio_catchup_test, audio_sink_resume_test, mod_packages_test, builtin_mod_manifests_test and host_enhancements_test pass. The package test's sandbox file-replacement denial was rerun successfully outside the sandbox in an isolated temporary directory.
• Title activation/catalog tests and real-package persistence check pass: fresh70, saved80 retained, upper180 accepted, reset70 retained, next-session activation and OFF fallback.
• Earlier continuity tests found ordinary stereo playback bit-identical at 44.1/48 kHz. The owner confirmed the continuity implementation removed clicks and personally found 70 ms stable in tested EU1 sections. This is not a full-campaign or all-platform certification.

Limits

70 ms is a tested recommendation on one system, not a universal minimum. Lower targets can stutter/drop out; 20 ms was unstable in the owner's tests. Host queue fill is not measured end-to-end latency: Windows/HDMI/eARC/speaker delay remains unmeasured. Long producer/display stalls can still cause recovery silence. No further audio optimization is proposed.

Source only, developed with AI assistance. No game data, credentials, private policies, unrelated mods or binaries.