fix: sync pr_state for stale-closed PRs excluded from scoring

[ai-mountain]

Summary

Fixes phantom OPEN rows on miner dashboards caused by the 35-day creation cliff introduced in #406. Stale-closed PRs now have their pr_state persisted while remaining excluded from credibility scoring — #406's intent (no fresh credibility penalty when miners close old backlog PRs) is preserved.

• Add stale_closed_pull_requests bucket on MinerEvaluation for CLOSED PRs that are observed by the scanner but excluded from scoring.
• Route the created_dt < lookback_date_filter skip in try_add_open_or_closed_pr into the new bucket instead of dropping the PR entirely.
• Persist the new bucket through the existing store_pull_requests_bulk / BULK_UPSERT_PULL_REQUESTS path — no schema change, no migration.
  Previously frozen pr_state='OPEN' rows now UPSERT to their real GitHub state.
• total_{merged,open,closed}_prs properties are intentionally unchanged — the new bucket is storage-only and does not inflate counts. This keeps calculate_credibility inputs and all scoring buckets identical to test. The dashboard mismatch (e.g. 7 Open in the table vs 3 in the header tile) resolves because the table now reads the refreshed pr_state='CLOSED', not because totals changed.

UPSERT semantics

BULK_UPSERT_PULL_REQUESTS zeroes scoring columns (earned_score, token_score, collateral_score, multipliers) on conflict. For a PR that was previously OPEN with collateral and is now CLOSED-not-merged and out-of-window, zero is the correct value — the PR no longer provides collateral. No partial-update query needed.

Related Issues

Fixes #768

Type of Change

• Bug fix
• New feature
• Refactor
• Documentation
• Other

Testing

• tests/validator/test_unscored_stale_prs.py — 5 tests covering: stale routing to the storage-only bucket, totals NOT inflated, fresh-PR regression guard, mixed fresh/stale, multi-stale totals invariant.
• tests/validator/utils/test_storage_mirror.py — 1 test confirming store_evaluation calls store_pull_requests_bulk with the new bucket as a separate (4th) call.
• Full validator test suite passes locally (708/708).
• Ruff lint + format clean.
• Pyright: 0 errors.

Checklist

• Code follows project style guidelines (ruff + pyright clean).
• Self-review completed.
• No schema change — reuses pull_requests table and BULK_UPSERT_PULL_REQUESTS query as-is.
• calculate_credibility inputs unchanged; fix: ignore stale closed PRs in lookback window #406's no-fresh-credibility-penalty guarantee holds.
• No change to total_*_prs properties; totals on /miners/{id} remain consistent with scoring buckets.

[ai-mountain]

@anderdc can you review this PR?

[ai-mountain]

@anderdc Thanks for merging this PR.