Improve speech dictation reliability

[jefking]

Summary

• add speech transcription audio stats, empty-transcript reasons, and structured error codes
• stop forcing browser dictation to language=auto so configured language hints are used by default
• add browser-side microphone signal checks and align compose defaults to base-int8/en

Validation

• go test ./internal/web
• go test ./...
• docker compose -f docker-compose.yml config --quiet
• synthetic speech PCM against local source server returned text
• synthetic silence PCM returned reason=empty_transcript with near_silent=true
• Playwright fake microphone smoke appended dictated text into the builder prompt