fix: isolate mirror PR cache entries

[Helios531]

Summary

Fixes a cache isolation gap where mirror PR objects returned from MinerEvaluationCache.get() could share ScoredMirrorPR instances with the cached snapshot.

When cache fallback is used, finalize_miner_scores() mutates top-level PR scoring fields such as earned_score, collateral_score, open_pr_spam_multiplier, credibility_multiplier, pioneer_rank, and pioneer_dividend. Legacy PRs were already protected by fresh top-level PullRequest copies, but mirror PR buckets were not.

Changes

• Add top-level copy isolation for:
  • mirror_merged_prs
  • mirror_open_prs
  • mirror_closed_prs
• Keep nested mirror metadata shared because the fallback path reads it but does not mutate it
• Preserve the existing cache-store behavior that strips ScoredMirrorPR.files
• Add regression coverage for:
  • mirror merged PR scoring-field mutations
  • mirror open PR collateral-score mutations
  • legacy PR top-level isolation
  • mirror file stripping at store time

Validation

uv run --with pytest pytest tests/validator/test_validator_cache_fallback.py](fix: isolate mirror PR cache entries)

Fixes #927

[anderdc]

Verified the claimed leakage doesn't manifest — every field listed in the issue is fresh-assigned in Phase 1 of finalize_miner_scores (scoring.py:319, 344-348) before Phase 2/3 read it. calculate_final_earned_score (scored_pr.py:94) recomputes earned_score from base × multipliers, so stale cached values get overwritten before observation. The cache also lives entirely in the GitHub-PR-fetch fallback path that's being decommissioned with the move to mirror-only scoring (classes.py:715). Closing as defensive isolation in a path on the way out.