fix: align parity evidence with BlindBench ingest

[nmogil]

Summary

Align multi-attempt provider-parity evidence with BlindBench's strict full-span ingest contract, and allow the completed paid matrix to be re-exported offline without rerunning providers.

• Give every strict evidence record a deterministic, attempt-specific run.id while retaining the real attempt ID and private logical grouping metadata.
• Exclude generated Python cache files from future trusted workspace evidence.
• Narrowly sanitize the serialized letter + colon + newline sequence that BlindBench's Windows-path detector treats as an absolute path.
• Add rollback-safe aggregate evidence re-export from retained checksummed bundles.
• Require independently unique run IDs and attempts in producer/parity validation.

Activation finding

The completed 18-run matrix was valid at the execution layer but could not be imported as one BlindBench batch because:

1. Mogil repeated stable run IDs across attempts while BlindBench requires independent uniqueness.
2. Three inventory projections triggered BlindBench's serialized absolute-path detector on source strings such as except ValueError as e:\n.

This change regenerates only reviewer-safe aggregate evidence. It does not alter retained bundle bytes or rerun any model/provider.

Verification

• pytest -q tests/test_harbor_bundle.py -k 'reexport or aggregate_pair or python_cache' — 5 passed
• pytest -q -m 'not daytona_smoke' — 120 passed, 1 paid Daytona test deselected
• ruff check . — passed
• mypy — passed
• git diff --check — passed
• Real retained run re-export — 18 valid JSON + 18 valid JSONL records
• BlindBench parseHarborEvidenceV1 — 18/18 parsed, 18 unique run IDs, 18 unique attempts
• All 18 retained bundle checksums remain valid
• Reviewer projection provenance/credential/canary/path scan — clean

Replacement safety

Both aggregate files are staged and validated before replacement. Each same-filesystem rename is atomic. An observed replacement failure triggers rollback of both prior destination states; this is deliberately not described as a crash-atomic multi-file transaction.

Safety boundaries

• No provider calls or paid execution
• No BlindBench upload
• No bundle mutation
• No training approval or export