Add Gemini support, AI summarization, provider selector, and integration test overhaul

[dmitry-kostin]

Summary

• Gemini multi-provider transcription — shared clients.py with is_gemini_model() + get_client(), OpenAI-compatible endpoint, prompt-driven diarization with A: B: labels
• AI summarization — Chat Completions API (OpenAI + Gemini), summary sidecar files ({id}_summary.md), Transcript/Summary tab toggle, copy per active tab
• Provider selector — /api/providers endpoint, frontend widget with localStorage persistence, per-provider model defaults in config
• Duration limit — "First N min" toggle, prepare_chunks() truncates audio via ffmpeg before chunking, 0–480 min validation
• Integration test overhaul — audio caching (in-memory + disk), LLM judge with confidence scoring (Gemini), debug reports with full outputs and scores
• Gemini prompt tuning — plain text for non-diarize, A:/B: labels for diarize mode
• json-repair — robust LLM JSON parsing for integration test judge responses

Test plan

• poetry run pytest — 188 passed, 8 deselected
• poetry run pytest -m integration — requires API keys + internet
• Manual test: transcribe with OpenAI and Gemini providers
• Manual test: summarize a transcript, verify sidecar file and tab toggle
• Manual test: provider selector switches models correctly
• Manual test: duration limit truncates audio

🤖 Generated with Claude Code