fix: strip mirror file contents in MinerEvaluationCache

[carlos4s]

Summary

MinerEvaluationCache got shallow-copy helpers in #786, but they only cover the legacy PullRequest lists. The three mirror lists (mirror_merged_prs, mirror_open_prs, mirror_closed_prs) still get shared by reference, which causes two real problems:

• Memory: cached evaluations keep ScoredMirrorPR.files[*].head_content and base_content (whole head/base file blobs). Cache memory grows with every mirror-scored PR and never shrinks ([Bug]:MinerEvaluationCache doesn't strip mirror file contents — unbounded memory growth #823).
• Aliasing: mutations on the source eval after store() leak into the cache, and back-to-back get() calls hand back the same ScoredMirrorPR / MirrorPullRequest / MirrorLinkedIssue / MirrorLabel objects.

This PR extends the same shallow-copy strategy from #786 to mirror PRs, via four small private helpers that mirror the existing _pr_for_cache / _pr_with_fresh_issues pair:

• _build_cache_entry: drop ScoredMirrorPR.files and shallow-copy the nested MirrorPullRequest (review_summary, labels, linked_issues + each linked issue's labels).
• _isolate_for_downstream: same nested-metadata copy on the way out, so successive get() calls return fully independent snapshots.

The cache-fallback path only runs issue competitions and issue discovery scoring, neither of which reads file contents — so dropping files outright is enough, no need to keep stripped MirrorFile placeholders.

Related Issues

Closes #823

Type of Change

• Bug fix
• New feature
• Refactor
• Documentation
• Other (describe below)

Testing

• Tests added/updated
• Manually tested

Three new tests in tests/validator/test_validator_cache_fallback.py::TestMirrorCacheIsolation, one per invariant:

• test_cache_drops_mirror_files_without_mutating_source — files is None on the cached copy across all three mirror buckets, and the source eval is untouched.
• test_store_isolates_cached_mirror_metadata_from_source_mutations — mutating source.mirror_merged_prs[i].pr.{title, review_summary, labels, linked_issues} after store() does not affect the cached copy.
• test_get_returns_isolated_mirror_metadata — two consecutive get() calls return independent snapshots.

$ uv run pytest tests/validator/test_validator_cache_fallback.py -v
9 passed

$ uv run pytest tests/
all green

$ uv run ruff check && uv run ruff format --check
clean

$ uv run pyright gittensor/classes.py tests/validator/test_validator_cache_fallback.py
0 errors

Checklist

• Code follows project style guidelines
• Self-review completed
• Changes are documented (if applicable)

[carlos4s]

@anderdc Please review this PR. I'd appreciate any feedbacks. Thanks

[anderdc]

Scope this down to the actual leak. The file-strip change in _build_cache_entry plus a _scored_mirror_pr_for_cache helper that nulls files is all #823 asks for.

Drop _scored_mirror_pr_with_fresh_pr, _mirror_pr_with_fresh_metadata, _mirror_linked_issue_with_fresh_labels, the _isolate_for_downstream mirror branch, and the two metadata-isolation tests (test_store_isolates_cached_mirror_metadata_from_source_mutations, test_get_returns_isolated_mirror_metadata) — those guard against mutations no caller on the cache-fallback path actually makes. The legacy _pr_with_fresh_issues deliberately only deep-copies issues for the same reason; see the comment at classes.py:723-725 ("PR metadata is read-only on the cache-fallback path, so we can share it"). The mirror path feeds the same downstream (issue competitions + issue discovery), so the same logic applies.

Keep test_cache_drops_mirror_files_without_mutating_source — that one matches the actual issue invariant.

If aliasing on metadata ever becomes a real concern, land it in its own PR with the offending caller cited.

[carlos4s]

@anderdc pushed the update. plz review again.