feat: add topic-specific clip selection

[seanphan]

Summary

• Add av clip selection from existing transcript and vision sidecars without re-ingestion.
• Keep timing source-aligned, enforce duration/deduplication/request-cap rules, and provide optional FFmpeg export receipts.
• Add a frozen synthetic evaluation corpus, independent labels, deterministic and labeled-mock comparison arms, held-out cases, and a reproducible cookbook.

Verification

• uv run pytest tests/ -q — 312 passed
• Ruff over all changed Python files — clean
• uv build — passed
• uv run av --help and uv run av clip --help — passed
• Offline dev/held-out receipts regenerated; all 30 rendered MP4s passed ffprobe

Evidence limits

The evaluation uses synthetic rights-cleared media and a labeled mock decision client. No live Jev allowance was available, so this PR does not claim live Jev quality, timing, cost, or provider parity. The two-second/two-cent social claim remains unverified.

[seanphan]

Supervisor verification at head 2d9d8a3ea217b66959723c9a9911e00194a389a6:\n\n- uv run pytest tests/ -q: 312 passed.\n- Ruff over all changed Python files: clean; uv build: passed; CLI help checks passed.\n- Offline synthetic dev/held-out receipts regenerated; 30 rendered MP4s passed ffprobe.\n\nEvidence is synthetic/labeled-mock only. No live Jev allowance was available, so this review does not claim live provider quality, timing, cost, or parity.