Speed up git-ai stats range mode by batching authorship-note reads

[jwiegley]

Summary

Makes git-ai stats <a>..<b> (range mode) faster — ~1.7× on a 100-commit range up to ~2.5× on a 1000-commit range — by removing redundant git-subprocess fan-out in the blame path. range_stats output is byte-identical.

Profiling (release build, real repo) showed the cost is git subprocesses, not SQLite — the default GitNotes backend touches 0 ms of SQLite on this path, so a #1630-style index would not help git-ai stats. The range path blames every changed file three times (diff_ai_accepted_stats + start/end VirtualAttributions), and each blame re-read every blamed commit's authorship note with a separate git notes --ref=ai show — thousands of per-commit note spawns, uncached across files and passes.

Changes (both output-preserving)

1. Batch the note reads. New notes_api::read_authorship_v3_batch resolves many commits' notes with one git ls-tree + cat-file --batch (via refs::notes_for_commits) instead of one git notes show per commit. It returns exactly what read_authorship_v3(repo, sha).ok() yields per commit — the v3 parse is factored into refs::parse_reference_as_authorship_log_v3 and shared by both paths, and the Http backend delegates per commit to preserve its cache-hit semantics. Both blame overlays (overlay_ai_authorship, populate_ai_human_authors) pre-seed their existing per-commit cache from one batched read.
2. Skip dead work on the stats path. blame() with no_output discards the blame hunks, so the populate_ai_human_authors pass that annotates them with ai_human_author (a second per-commit note read) is wasted there. Adds GitAiBlameOptions::skip_human_author_population (default false, debug-asserted to only be set on no_output paths) and sets it on the two stats blame call sites. It is output-invariant because overlay_ai_authorship derives line_authors per line independently of hunk grouping.

Performance (release build, real repo, same-run A/B, range_stats byte-identical)

The speedup grows with range size because it eliminates per-commit note fan-out, which scales with the number of blamed commits:

command	baseline (same run)	optimized	speedup
git-ai stats HEAD~100..HEAD	69.0 s (noisy: 69–93 s across runs)	40.0 s	~1.7×
git-ai stats HEAD~1000..HEAD	272.3 s	110.2 s	~2.5×

Under a subprocess-counting harness on HEAD~100..HEAD, per-commit git notes show collapses from ~7,200 to ~50 (the remainder shifts into batched cat-file/ls-tree); total git spawns ~14.5k → ~9.3k same-harness. (Exact spawn totals are method-sensitive; the notes show collapse and the wall-time deltas are the robust figures.)

Validation

• task fmt, task lint (clippy -D warnings), full task test — green.
• Same-run A/B byte-identical check of range_stats on real data across 100- and 1000-commit ranges (range_1000 = 120,390 AI lines across 32 tool/model combos — identical before/after).
• Regression tests: batched read ≡ per-commit read (commits with notes, without notes, non-existent); empty-input fast path; and skip-dead-pass gate output-invariance (blame line authors identical with the pass on vs off).
• Two independent adversarial reviews (the batching, and the skip-dead-pass gate) — no blockers. Documented parity boundary: on GitNotes the batched read decodes note blobs with from_utf8_lossy (matching the existing CommitAuthorship path) vs the per-commit strict decode — unreachable for real notes (always valid UTF-8 JSON). The Http and batch-error fallback arms delegate per commit to the existing reader (equivalent by construction; not exercised by tests).

Note

The #1630 index pattern still applies — just to a different command, git-ai usage (MetricsDatabase::get_metric_history is a full-table scan that deserializes every row). Not touched here.

🤖 Generated with Claude Code

[devin-ai-integration]

✅ Devin Review: No Issues Found

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

[jwiegley]

• Speed up git-ai stats range mode by batching authorship-note reads #1652 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.