[R44VC0RP]

"While dictating" gains Muffle behind Labs › Muffle other audio (labs.muffle, off by default, macOS): other apps keep playing, but low and through a wall, like a club heard from its bathroom.

How it works

• A muted-when-tapped global Core Audio process tap (Hex's own tones excluded) silences other apps at the output; a private aggregate device holding the tap and the default output replays them through the signal path from its IOProc. Needs System Audio Recording (macOS 14.2+).
• Signal path (src/audio/muffle.rs, pure, every OS): 400 Hz RBJ low-pass that shuts with a Q 3 → 0.707 thump over 150 ms, level ramped to 40 % over 80 ms, a small Freeverb-shaped room (4 combs, 2 allpasses, +23 samples on the right, feedback 0.72, damp 0.45, wet 0.0875), tanh soft clip above 0.9. Opening sweeps the cutoff to 20 kHz with Q 4 → 0.707 over 200 ms before the tap is torn down.
• Effect::Muffle composes with the other held effects in recording_environment. Without the Labs feature or the permission, or if the tap fails, the session lowers the audio instead; Linux and Windows treat a saved muffle as Lower audio.
• General shows the Screen & System Audio Recording row under Muffle (shared with Also pause the browser); the Labs toggle asks for the permission when turned on, never while dictating.
• hex settings: recording_audio_behavior accepts muffle; labs.muffle added.
• meeting::create_tap_aggregate takes an optional output sub-device (the muffle aggregate uses it to play; existing callers pass None).

Known gap: a default-output change mid-dictation keeps the old output until the session ends; dictations are short, so no rebuild yet (noted in docs/features/labs.md).

Verification

• scripts/check: fmt, code-shape, clippy, 2208 tests pass.
• audio::muffle unit tests: 4 kHz vanishes and 80 Hz passes at the muffled level once shut; opening restores the dry signal within 2 %; silence stays silent and loud input stays bounded; the sweeps start and end where they should.
• Ignored on-hardware test muffles_the_default_output_for_a_while ran on a Mac with a test tone playing: 271 IO cycles in 3 s, tapped peak 0.60, the door opened before teardown.
• Goldens: the Labs page gains a row, so the Linux golden tests need regenerating (check-linux.yml → golden-actual).