Changed Prompt and added feedback voice

[Butenkite]

Summary

After LLM feedback is generated for a per-question scorecard or an end-of-interview session review, the selected interviewer voice reads that feedback aloud via the existing `/speech/tts` endpoint. Long feedback scripts are split into Groq Orpheus\u2019s 200-character chunks and played sequentially. The LLM system prompts for `/feedback/generate` and `/feedback/generate-session` now require second-person, in-the-moment coaching language instead of third-person candidate evaluation phrasing.

Feedback and speech flow

Flow	Trigger	Spoken content	UI behavior
Per-question feedback (new)	`ScorecardPanel` `onFeedbackReady` after `POST /feedback/generate`	Script from `buildQualitativeFeedbackScript` (summary, strengths, improvements, delivery notes)	Auto-opens scorecard when stage is `done`; disables Next Question while audio plays; status shows interviewer sharing feedback
End-of-interview feedback (new)	`SessionScorecardPanel` `onFeedbackReady` after `POST /feedback/generate-session`	Script from `buildSessionFeedbackScript` (overall summary, strengths, improvements, delivery notes)	Speaks once per finished session; status updated during playback
TTS chunking (new)	Any `speakWithGroq` call with text over 200 characters	Sentence/word-boundary chunks sent to `POST /speech/tts` one at a time	Prevents TTS 502 failures on long feedback text
LLM voice/tone (modified)	`/feedback/generate` and `/feedback/generate-session`	Shared `_VOICE_INSTRUCTIONS` block in system prompts	Feedback fields address the user directly with conversational coaching language

File-by-file changes

Backend / API

• `backend/services/llm/router.py`: Adds shared `_VOICE_INSTRUCTIONS` and rewrites per-answer and session system prompts to require direct second-person coaching tone so generated feedback reads naturally when spoken aloud.

Frontend components

• `frontend/app/InterviewClient.tsx`: Wires feedback TTS after scorecard fetch, auto-shows per-question feedback on `done`, tracks speaking state, aborts playback on dismiss/next question, and imports shared TTS/speech-script helpers instead of an inline TTS implementation.
• `frontend/app/scorecard/ScorecardPanel.tsx`: Adds optional `onFeedbackReady` callback invoked when fetched or pre-resolved feedback is ready so the interview client can trigger spoken playback.
• `frontend/app/scorecard/SessionScorecardPanel.tsx`: Adds optional `onFeedbackReady` callback invoked when session feedback fetch completes so end-of-interview spoken feedback can start.

Shared / utilities

• `frontend/lib/interview-coach/feedbackSpeech.ts`: New helpers that assemble per-question and session LLM feedback JSON into a single spoken script for TTS without generating new content.
• `frontend/lib/speech/tts.ts`: New shared `speakWithGroq` with 200-character chunk splitting, sequential playback, and abort support, replacing the prior inline implementation in `InterviewClient.tsx`.