feat(irl-source): video delay follows the audio playout offset

CLAUDE.md

@@ -69,8 +69,8 @@ scripts/package.sh linux target/release dist       # the release archive, locall
 - `irl-core` and `ffmpeg` have unit tests in each module. Everywhere else, tests live in `tests/`. The link arguments that resolve libobs only reach integration-test targets, so `crates/irl-source` sets `test = false` on its lib and `crates/obs` keeps its lib free of `#[cfg(test)]`. Shared fixtures for irl-source live in `crates/irl-source/tests/common/`.
 - Tests that touch libobs run on Linux. CI skips them on Windows and macOS. `calldata_*` is safe anywhere. On a Mac, `make test-shim` runs the pacing, audio-core and network tests against a four-function stand-in; plain `cargo test -p irl-source` there dies with SIGSEGV before its first assertion.
 - `tests/network_sim.rs` drives the real jitter buffer, PTS repair, speed controller, output clock and packet queue against a synthetic sender on a virtual clock (stall, burst, dropouts, a sender off wall clock, video behind audio). It asserts the design's promises: the OBS clock never jumps except across a declared restart, audio is never skipped once primed, latency does not ratchet, decoded memory does not grow with Target Buffer.
+- `tests/av_sync_sim.rs` is the end-to-end A/V sync sim: it measures the real offset between picture and sound as libobs is handed them. Run it whenever you touch the video delay, the audio hold or the playout mapping. `docs/skew-testing.md` reproduces the same cases on a real OBS with `crates/irl-core/examples/skew-relay.rs`.
 - Where you read the buffer fill decides the number. It oscillates by one chunk within every cycle: before the pump's read it averages the target, after it a chunk lower. The stats line's `buf=` is a random sample of that oscillation.
-- `tests/video_pipeline.rs` reads the real clock and can flake under parallel load; rerun with `-- --test-threads=1` before suspecting the code.
 - Real-stream validation is manual: run the same feed through this build and a known-good one and compare the 30-second stats line field by field.
 
 ## CI and releases