perf(record): keep one speech model loaded — 2948 ms → 77 ms to start recording

[meocong]

Measured on the released build, gipformer-65m, from session_start to session started:

ready before warming: null
cold  : 2948 ms to start recording
warming took 3024 ms → {"model":"gipformer-65m","language":null}
warmed:   77 ms to start recording
ready again after the session: {"model":"gipformer-65m"}      ← refilled by itself

Nearly three seconds, paid at the start of every meeting, with nothing on screen explaining the
pause. Not disk — a second construction in the same process costs the same as the first — it is
ONNX Runtime building the session.

The shape

• One decoder, not a cache. A second warm model doubles resident memory (~150 MB idle) for the
  rarer case of switching language between meetings.
• One lane. A two-lane session still builds its second decoder; two lanes cannot share mutable
  inference state. All of the common case is instant.
• Given away, not lent. A session owns what it takes; the slot refills afterwards. No question
  about a killed recording holding a borrowed decoder.
• A miss, never a substitute. A decoder built for another language is not reused — silently
  transcribing Japanese with a Vietnamese decoder is the failure this whole area exists to prevent.
• Removing a model clears the slot: a warm decoder pointing at deleted blobs is a crash waiting for
  the next recording.

Also fixed

model_load — the command whose comment says "so the first recording is not delayed by it" — was
answering cannot load: this binary was built without recognition support in builds that have
it, the same defect model_swap had until yesterday. Both arms now work.

/status gains ready beside the existing fields (not wrapped around them, so no client breaks),
and the record card says "Sẵn sàng ghi ngay". Warming with nothing installed answers 200 with a
reason — the browser suites caught it logging a 400 on a first run where nothing was wrong.

Verified

1289 Rust tests (4 new in warm.rs), 297 web tests, shell / language / full-flow suites,
clippy -D warnings on both feature sets, fmt, tsc, eslint.

🤖 Generated with Claude Code