akou 0.6.3

0.6.3 — a server that hands work to other akou servers keeps working itself while they are busy

On 0.6.2 a job that a server.remotes entry names waited whenever every remote that offers it was busy, even when the primary could run it and sat idle. A primary that is a good worker itself could not share a backlog with its remotes: it either kept every job or watched them queue behind its slower helpers. In 0.6.3 one setting lets it run those jobs itself while its remotes are full.

Server

• server.remotes_overflow. Off by default, so nothing changes until you turn it on. On: a job a remote entry names runs on the primary while every remote that would take it already holds its jobs, instead of waiting. A job only a remote can run still waits for one, and after a start a job waits until the remotes it names have answered their first probe. Turning it on over the API routes the waiting jobs at once. Settings shows it under Other servers (#340).

Behind the scenes

• The macOS nightly no longer dies in the Qwen stage on a runner short of memory (#339).
• A 1.0.0 release is refused while any gate has only half a Pass on record (#341).

Known limitations

• A remote that goes to sleep with a job keeps its copy and may finish it after the primary has run the job again elsewhere. Only the primary's answer reaches the client.
• Every item under 0.6.2'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

• fix(nightly): stop the macOS nightly dying in the Qwen stage on a runner short of memory by @GeiserX in #339
• fix(release): refuse 1.0.0 while a gate has only half a Pass by @GeiserX in #341
• feat(server): let a primary run a remote-named job itself while every remote is busy by @GeiserX in #340
• docs(gates): re-measure containment and cold start on 36a2237 so 1.0 does not rest on September numbers by @GeiserX in #347
• fix(compose): keep the Telegram-Archive example on the released akou, and warn that public_host turns away service-name calls by @GeiserX in #345
• docs(gates): record G7 on the signed 0.6.2 app, so the harness gate has a verdict before 1.0.0 by @GeiserX in #344
• chore(release): 0.6.3, so a server that hands work to other akou servers keeps working itself while they are busy by @GeiserX in #348

Full Changelog: v0.6.2...v0.6.3