perf+fix: diff-based attribution, unified rebase path, daemon guard fixes

[svarlamov]

Summary

Addresses review feedback from #813 and adds significant performance, correctness, and simplification improvements to the rebase authorship rewriting pipeline.

Performance: diff-based line attribution transfer

• Replace the expensive char-level byte diffing (AttributionTracker::update_attributions) and VirtualAttributions wrapper with lightweight line-level diffing using the existing imara-diff Myers algorithm
• Eliminates both VA construction overhead and O(n*m) char-level transform, replacing with O(n+m) line-level positional transfer
• ~4x improvement for per-commit transform, ~3.4x for full pipeline
• Fixes the duplicate-line collision bug in content-matching (positional tracking instead of HashMap<line_text, author>)

Performance: fast cached serialization

• Restore per-file attestation text caching and metadata JSON template pre-splitting
• Assembly is pure string concatenation instead of full serialize_to_string() per commit
• Eliminates serialization as a bottleneck in the commit processing loop

Fix: prevent pipe deadlock in exec_git_stdin

• Writing all stdin data before reading stdout causes deadlock when the child process's stdout pipe buffer fills (~64KB)
• Fixed by spawning stdin writes in a separate thread, allowing concurrent read/write

Fix: diff-tree --stdin parsing for -z null-terminated output

• With -z flag, commit SHA headers in git diff-tree --stdin output are null-terminated, not newline-terminated
• Parser was silently returning 0 changed files, disabling the parallel blob loading optimization

Refactor: unified rebase path

• Removed the use_line_lookup_fast_path branching that maintained two parallel transform and serialization paths
• Both paths already used identical diff-based transfer; the only difference was serialization (cached string fragments vs structured AuthorshipLog)
• Unified to the structured AuthorshipLog approach, eliminating a class of divergence bugs
• Net -230 lines of code removed

Fix: per-commit prompt metrics

• Prompt metrics (accepted_lines, overridden_lines) now update per commit instead of being frozen from initial state

Fix: file deletion/reappearance inheritance

• Preserve attributions and file contents through file deletion so files recreated later in the rebase sequence inherit via positional diff transfer

Fix: daemon event handler guards

• Add is_valid_oid/is_zero_oid guards to RefUpdated handler (matching Reset handler)
• Add missing is_ancestor_commit guard to RefUpdated handler, preventing spurious rebase events on backward ref updates
• Revert first_parent_oid onto_head hint — incorrect for multi-commit rebases
• Swap reflog fallback order in parse_update_ref_heads (target ref before HEAD)
• Skip rebase-like reset handling for backward resets

Other

• Support update-ref and rebase-like reset detection in daemon mode for Graphite compatibility
• Unignore 5 Graphite tests that now pass via update-ref hook
• Keep pretty-printed JSON in authorship notes metadata
• Clean up dead code (upsert_file_attestation, unused timing variables)

Test plan

• 76 rebase integration tests pass (0 failures)
• 21 Graphite integration tests pass
• 24 rebase unit tests pass
• 5 new unit tests for diff-based transfer (equal, insert, delete, replace, duplicate lines)
• 2 new regression tests: per-commit prompt metrics, file delete-recreate attribution
• 2 benchmark tests (ignored): full pipeline comparison, scaling 50-5000 lines
• Lint & Format fully green
• Ubuntu tests fully green (all 3 modes + SCM e2e)
• Git Core Compatibility green
• Performance Benchmarks green

🤖 Generated with Claude Code

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 2 additional findings.