perf(recording): wake the audio mixer on input instead of polling every 5 ms

[richiemcilroy]

What

The audio mixer's run loop slept a fixed 5 ms between ticks for the whole recording, about 200 wakeups/s. Most ticks found nothing new.

Now:

• Input wakes the mixer. Input channels are polled with a waker that unparks the mixer thread (poll_next_unpin instead of try_next; the same empty/ready/closed semantics), so a producer's send wakes it straight away.
• Idle wait. With no input, it waits park_timeout(idle_wait()): half the shortest source buffer timeout, clamped to 5–25 ms (10 ms for a wired mic's 20 ms minimum). A source that goes quiet is still filled with silence, and stalls are still observed, within half its timeout.
• Unchanged: output placement stays sample-based, so tick timing doesn't move audio. The stop/drain path (finish) is untouched.

Expected effect: the mixer ticks per input frame (tens per second for a mic, ~100 for system audio) plus the occasional idle timeout, instead of a fixed 200/s, and input is processed up to 5 ms sooner.

Proof

• Unit tests:
  • input_wake::a_frame_arriving_after_a_drain_wakes_the_mixer
  • input_wake::idle_wait_is_half_the_shortest_buffer_timeout_within_bounds
• A/V sync and real-capture A/B against main: the Verify recording dispatch runs are linked in the comments below.

No Hona-authored lines are touched.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
cap-status	Ready	Preview	Oct 20, 2026 12:39pm UTC

[greptile-apps]

Greptile has paused reviews on this repository — it used its 1000 free open-source review credits for this billing period. Reviews resume automatically on October 26. To continue before then, an organization admin can keep reviews running past the free credits — those bill as normal usage.

[capy-ai]

Capy found no issues.

View 2 other findings in Capy.