feat(api/transcription): include segments + duration + language on stream done event

[localai-bot]

Summary

Extends the transcript.text.done SSE payload emitted by streamTranscription to additively carry language, duration, and a segments array so streaming clients can build the same TranscriptionResultSeconds shape they get from the non-streaming JSON path.

Why

Streaming clients (e.g. notetaker / notary) want streaming semantics for one big practical reason: the JSON path's ResponseHeaderTimeout trips when whisper requests queue behind each other on a SingleThread backend. streamTranscription fixes that by flushing 200 + headers immediately, but until now the done event only carried text — clients that needed per-utterance timings or audio duration had to fall back to the JSON path and accept the queue-induced timeouts again.

This change closes the loop: the streaming endpoint becomes a strict superset of the JSON endpoint's body, and clients can stay on SSE end-to-end.

Compatibility

• OpenAI streaming spec: only text is mandated on transcript.text.done; spec-compliant clients ignore unknown fields.
• Empty values are still omitted (Language == "", Duration == 0, len(Segments) == 0), so a transcription that came from a backend without these signals emits the same shape as before.
• Wire format: start/end are float seconds (matching TranscriptionSegmentSeconds.Seconds() from the JSON path) so clients can share decoding logic.

Test plan

• Manual: hit POST /v1/audio/transcriptions with stream=true against a model that emits segment timings (whisper-cpp). Verify the final data: {...} line before [DONE] includes segments and duration and matches the JSON-mode response for the same audio.
• Manual: same request against a backend that doesn't produce segments. Verify the done event still includes text and the segments field is absent (omitempty preserved).
• Existing non-streaming JSON path is untouched — no regression risk for spec-compliant clients.

Companion change

The notetaker repo's notary client switches its Transcribe to streaming and consumes these fields in the same shape as its existing JSON-path parser (internal/shared/localai/client.go::parseTranscriptionStream). Without this server change, the streaming client falls through to text-only results — so this PR completes the round trip but isn't a breaking dependency.

🤖 Generated with Claude Code