feat(suggestions): skip live suggestions for backchannel interviewer turns

[alpha5611331]

Summary

• Adds classifyInterviewerTurn() (src/main/utils/interviewer-turn.ts), an in-process, deterministic first stage that runs on the merged interviewer turn and returns Skip (pure backchannel - drop outright, no request, no card), Answer (a completed question or directive - generate immediately), or Uncertain (park on an INTERVIEWER_TURN_SETTLE_MS timer that any further interviewer final re-arms, so a question an ASR final split mid-sentence is judged whole).
• Skip only fires when a backchannel lexicon consumes the entire turn from the front - deliberately asymmetric, since a missed question is silent and unrecoverable mid-interview while a wasted request just costs one ignored card.
• transcript.service.ts now schedules the suggestion through this classifier instead of firing unconditionally on every interviewer final.
• Answer / Uncertain ride the request as turn_verdict for the backend's own speculative gate (PowerInterviewAI/backend#49).
• suggestion-live.service.ts no longer appends a Pending card the moment a request starts - it holds for LIVE_SUGGESTION_RENDER_DELAY_MS, so a turn the backend gate suppresses in that window produces no card at all instead of one that flashes and is retracted.

Fixes #100.

Test plan

• pnpm lint - clean
• pnpm exec tsc -p tsconfig.electron.json --noEmit / tsconfig.app.json - clean
• pnpm test:main - all checks passed
• New: test/interviewer-turn.test.mjs (Skip/Answer/Uncertain classification, both directions of the asymmetry, fragment-waits-for-continuation)
• Manual: run an interview, confirm "mhm"/"okay"/"got it" produce no card and a question split across two ASR finals still gets one coherent answer

[gitar-bot]

Code Review ✅ Approved 2 resolved / 2 findings

Adds deterministic classification to skip live suggestions for backchannel interviewer turns and delays pending card creation to prevent UI flashing. Addresses findings regarding block inspection scope and documentation links.

✅ 2 resolved

✅ Edge Case: Skip verdict only inspects the last merged interviewer block

📄 src/main/services/transcript.service.ts:108-121
scheduleSuggestion classifies only cleaned.filter(Other).slice(-1)[0] (transcript.service.ts:111), not the full set of interviewer transcripts since the candidate last spoke. If an interviewer turn is split into two separate Other blocks (a pause > TRANSCRIPT_INTER_TRANSCRIPT_GAP_MS = 5s), and the final block is pure backchannel, the whole turn is Skipped — dropping any unanswered question in the earlier block entirely (no request, no card). Within 5s the blocks merge and re-classification is whole, so this is a low-likelihood edge, but it contradicts the 'runs on the merged turn' claim for multi-block turns and is exactly the silent-drop the asymmetry is meant to prevent. Consider classifying the concatenation of trailing Other blocks (or only returning Skip when every interviewer block since the last Self turn is backchannel).

✅ Quality: CLAUDE.md links turn_verdict types to the wrong file

📄 CLAUDE.md:74 📄 src/main/types/llm.ts:72-83
CLAUDE.md line 74 states GenerateLiveSuggestionRequest / the turn_verdict wire values live in src/renderer/types/llm.ts ('types/llm.ts mirrors the wire values answer / uncertain'). Those types and the new RequestTurnVerdict enum are actually defined in src/main/types/llm.ts; the renderer file contains neither GenerateLiveSuggestionRequest nor RequestTurnVerdict. Update the doc link to point at src/main/types/llm.ts to avoid misleading readers.

Options

Auto-apply is off → Gitar will not commit updates to this branch.
Display: compact → Showing less information.

Comment with these commands to change the behavior for this request:

Auto-apply	Compact
gitar auto-apply:on	gitar display:verbose

_{Was this helpful? React with 👍 / 👎 | Gitar}