Generate concise summaries and full proof expositions

[tadamcz]

Summary

• generate a concise proof summary and a full natural-language proof or disproof independently for every accepted result
• replace the previous proof metadata fields with summary and full_proof objects containing text and generation provenance
• allow unrestricted full expositions, with 3,000 words explicitly described as acceptable when needed
• support subset=all for full-dataset result runs
• update deterministic collection and focused tests for the new schema

Validation

• uv run pytest -q tests/test_summarize.py
• uv run mypy scripts/summarize/task.py scripts/summarize/collect.py
• uv run ruff check scripts/summarize/task.py scripts/summarize/collect.py tests/test_summarize.py
• production run over all 15 published result runs: 829 accepted proofs, 1,658 successful samples, zero sample or schema audit errors

Full generated proofs ranged from 42 to 2,556 words.