Echo 1.1 - transcription actually works

Fixes the three bugs that made 1.0's transcripts fiction.

What was wrong. 3h12m of capture produced 22,218 words, almost none of them spoken: "the security is over" seven times in a row, and one line reading मार žе дусов लाहथ प buffalovia hurry up — Devanagari and Cyrillic in the same sentence.

Three causes, all fixed:

1. Whisper was fed the silence. A chunk that was 3% speech went in whole, and the model invents fluent text over room tone. Only voiced regions are now sent, with timestamps mapped back to real clock time.
2. Language auto-detection reads only the first 30 seconds and applies the result to the whole call — across 99 languages. A chunk opening with silence had its language chosen from noise. Detection is now restricted to en/hi/mr.
3. The JNI never enabled whisper's own defences: suppress_nst defaults to false upstream, and per-segment no-speech probability was never read.

Plus a per-language initial prompt (without it, Marathi came back romanised), level normalisation, and repetition collapse at both transcription and summary time.

Throughput. Base measured 714 s to transcribe a 256 s chunk — 0.4x realtime, a queue that never drains. The cause was the temperature-fallback ladder retrying every degenerate window up to six times; it is now capped at three.

Model guidance, measured rather than asserted:

model	Hindi	Marathi	code-switched	English
Base	0.23	0.00	0.08	1.00
Small	0.54	0.13	0.38	1.00

Small is the right choice for Devanagari, at roughly 2.5x the compute.

63 unit tests pass. See docs/VERIFICATION.md for what remains unverified.