Filter cosmetic orphan changes from semantic diff

[Iron-Ham]

Summary

• classify orphan changes with parser structural hashes so comment-only orphan edits are cosmetic
• keep real orphan code changes visible under --no-cosmetics and preserve their summary buckets
• treat shebang edits as structural executable metadata

Fixes #272

Test plan

• cargo test --workspace
• dummy Git repos covering comment-only edits, added comments, added imports, modified imports, and shebang edits
• independent subagent review completed; actionable findings fixed
• claude -p reviewer launched twice; final embedded-diff run exited successfully but produced no review text

[inspect-review]

inspect review

Triage: 38 entities analyzed | 0 critical, 0 high, 17 medium, 21 low
Verdict: standard_review

Findings (6)

1. [low] In recalculate_diff_summary, orphan changes are now double-counted in change-type buckets. When entity_type == "orphan", the code increments orphan_count but no longer returns early, so it falls through to the match statement and increments the change-type bucket again (e.g., added_count). This means orphans contribute to both orphan_count AND their respective change-type count, which breaks the invariant that orphan_count should be a subset cross-cutting metric.
2. [low] In recalculate_diff_summary, the logic for Moved/Renamed/Reordered changes double-counts modifications. When a change is Moved/Renamed/Reordered AND has content changes, it increments both the move/rename/reorder count AND modified_count. However, if the change is also an orphan, it will be triple-counted (orphan_count + move/rename/reorder count + modified_count), leading to incorrect summary totals.
3. [low] In retain_non_cosmetic_changes, the filter logic is inverted for compound changes. The code retains changes where structural_change != Some(false) OR the change is Moved/Renamed/Reordered. This means a cosmetic (structural_change == Some(false)) Moved/Renamed/Reordered change will be retained, which contradicts the intent to drop purely cosmetic changes.
4. [low] In test no_cosmetics_filter_recomputes_summary_from_remaining_changes, the expected modified_count changed from 1 to 2, but the test setup only has one Modified change (the structural one). The orphan has ChangeType::Modified but should not be counted separately in modified_count if the old behavior was correct, suggesting the new counting logic is wrong or the test expectation is incorrect.
5. [low] In recalculate_diff_summary, the test orphan_changes_count_toward_change_type_buckets expects orphans to count toward change-type buckets (asserts added_count == 1 and deleted_count == 2 with 1 orphan added and 2 entities deleted). However, the comment in the AFTER code says 'orphan_count is cross-cutting metadata, while retained orphans still contribute to change-type buckets', which contradicts the BEFORE comment that said 'Orphans only skip the entity change-type buckets'. The removal of the continue statement changes the behavior to make orphans count in both places, but this appears to be inconsistent with the original design where orphans were separate.
6. [low] In retain_non_cosmetic_changes, the filter logic now allows cosmetic changes (where structural_change == Some(false)) to be retained if they are Moved, Renamed, or Reordered. However, in recalculate_diff_summary, these cosmetic compound changes will increment both their primary bucket (moved_count, renamed_count, reordered_count) AND modified_count if has_content_change() returns true. This creates a logical inconsistency: a cosmetic change (structural_change=false) should not have a content change, but the code will still check and potentially double-count it.

---

Reviewed by inspect | Entity-level triage found 0 high-risk changes

[Iron-Ham]

Superseded by #311.