Partial transcription, in-place history rewrite, live UI polish

[etiennechabert]

Summary

Streaming-style transcription: while a speaker keeps talking, Polyglot re-runs Whisper + translation on the growing buffer every 10 s and emits a payload with the same utterance_id. The frontend upserts by utterance_id, so the same row updates in place; when the final batch flushes (speaker switch or 60 s cap) it replaces the partial with the polished version.

What ships

Backend

• partial_transcribe_and_emit() runs Whisper + translations on a buffer snapshot, uses _current_bot_speaker as the speaker label, skips diarization / transcript-file write / summarization accumulation.
• Shares transcription_lock with the final so partial + final Whisper calls serialise on the GPU.
• Mints a fresh utterance_id on the first chunk of every batch (in both the is_processing=True and =False branches of process_audio) and resets it when the final flushes, so every partial + final for one turn shares the same key.
• Final WS payloads now carry utterance_id + is_partial:false; partials carry utterance_id + is_partial:true.
• Admin auth replays the current bot_status (including the meeting URL) + meet_roster so a late-arriving admin tab reflects the active bot.
• Cache-Control: no-store headers on the viewer + admin routes so browsers always fetch fresh templates.

Bot

• Emits a bot_info event on connect with the meeting URL so the admin panel shows which call the bot is attached to, even when the bot was started from the CLI.
• nameFromTile() now strips the "'s Presentation" suffix so a screen-sharing participant's tile doesn't label them "X's Presentation".

Frontend (admin + viewer)

• Transcript is keyed by utterance_id. Partial arrivals upsert the matching row; finals replace it. Timestamps are preserved across the swap so rows don't reshuffle.
• Partials get a dashed left border + "● live" tag + 0.75 opacity; finals get a solid border + full opacity.
• Viewer row segments render in a vertical body column instead of the flex-row that was making multi-segment utterances render horizontally.
• Viewer keeps the ?p=<password> query param across refreshes so the user doesn't have to re-enter the passphrase on every reload.

Test plan

• Validated live against a real Meet call: partials fire every 10 s, carry the same utterance_id, final at 60 s cap flips the bubble from dashed to solid and mints a new id for the next turn.
• [BOT] Resolved SPEAKER_XX → Etienne Chabert's Presentation proves Phase 5 still works end-to-end on finals.
• Diagnostic socket.io client confirms partials reach admin + lang_en rooms when viewers are joined.
• Admin bot-status + URL replay on auth verified with a manual refresh after bot was already running.

🤖 Generated with Claude Code