akou 0.6.4

0.6.4 — a server whose remotes all drop at once dispatches again when they return

On 0.6.3 a primary with server.remotes could stop dispatching for good. When every remote timed out in the same moment while the 30-second probe was in flight, the probe compared the remotes against the state it had seen before its requests, saw no change, announced nothing, and the queue sat with hundreds of jobs queued, none running and every remote reported up, until a restart. In 0.6.4 the probe compares against the state after its requests, so a remote marked down mid-probe is announced up and the queue runs again.

Server

• Dispatch resumes after every remote drops at once. The probe takes its change snapshot after its two requests, so a remote.down written while they were in flight ends in remote.up and a dispatch, with the local worker's own jobs untouched (#353).

Behind the scenes

• The macOS nightly no longer dies when Bun's hidden 6-minute fetch limit cuts a slow Qwen request; the night's requests pass timeout: false and the macOS leg has 200 minutes (#352).
• G6 is recorded as Partial on fp32 with both machines measured: speed passes, the committed line misses 1.5 s on the call channel while the diarizer holds that line for its speaker (#350).

Known limitations

• A stalled queue still reports as healthy over the API (queued N, running 0, every remote up); a signal for it is a follow-up.
• A job only a remote can run that was refused with 409 waits for the next submit, probe change or remote job end before it is offered again.
• Every item under 0.6.3's Known limitations still applies.

Signed and notarized. The app and the akou command line are signed with a Developer ID and notarized by Apple, so macOS opens them without an extra step, and the microphone and system audio grants should survive updates. Coming from 0.6.0 or older, which were not signed, macOS asks for both grants once more. The speech models are downloaded on first run. See docs/getting-started.md.

What's Changed

• docs(gates): record G2 passing inside the packaged app on all three OSes by @GeiserX in #342
• fix(gates): make the macOS 14.3 VM check able to answer, and show the tap opens on 14.3 by @GeiserX in #346
• feat(gates): prove the window, tray, hotkey, login item and headless start on Windows and Linux by @GeiserX in #343
• docs(gates): G6 is Partial on fp32, because the call channel waits for its speaker past 1.5 s by @GeiserX in #350
• fix(nightly): stop Bun's hidden 6-minute fetch limit killing the macOS night's slow Qwen requests by @GeiserX in #352
• fix(server): resume dispatch after every remote drops at once instead of stalling the queue by @GeiserX in #353
• chore(release): 0.6.4, so a server whose remotes all drop at once dispatches again when they return by @GeiserX in #355

Full Changelog: v0.6.3...v0.6.4