tests: regression for scene-wrapped source audio (#1493)

[summeroff]

What

Adds an integration test reproducing #1493: an ffmpeg_source wrapped in a scene (with sceneItem.video deliberately left unset, so the scene item's canvas stays NULL) and set as the output source must still record non-silent audio.

Before the libobs fix this case recorded silence; with it, the source's audio is mixed through the scene as expected.

Why it caught the bug

The existing dual-output tests pass only because they assign sceneItem.video (which calls obs_sceneitem_set_canvas). The plain scene.add() path — and the reporter's repro — never set it, leaving the item canvas NULL. This test exercises exactly that path.

How it asserts

New tests/osn-tests/util/media_probe.ts reads mean_volume via ffmpeg's volumedetect filter (ffmpeg resolved from FFMPEG_PATH or PATH). Digital silence reports ~ -91 dB, real audio sits well above -40 dB, so the test asserts mean_volume > -80 dB. Helper logic validated locally against generated silent (-91 dB → fails) vs tone (-21 dB → passes) clips.

Dependency

Requires the libobs fix — streamlabs/obs-studio#738 — to be present in the consumed libobs build. This PR will not go green until that fix is rolled into the osn libobs version.

Notes

• Type-checks clean (tsc --noEmit, project options).
• Adds a new test-time dependency on an ffmpeg binary; OBS bundles one, or set FFMPEG_PATH.

🤖 Generated with Claude Code