Facts ingest: batch phase-B blob measuring, per-phase timing logs, and tighter ingest scope #862

Why

A cold pixel build-index --history . on a worktree of this repo took 253 s. Reproduced on a fresh clone of the same repo: 140.7 s, with 275 of 287 work ticks spent in phase B (blob measuring). Phase A (metadata) and phase C (diff text) are cheap; phase B is the bottleneck.

Evidence (measured 2026-10-06)

Root cause: measure_commit_blobs (crates/pixel-facts/src/ingest.rs:803) measures one git ls-tree -l process per changed path (blob_size, ingest.rs:847), plus a second spawn per rename against the parent tree. The repo has ~9,170 changed-path rows across 1,024 commits → 10k+ git spawns at ~7 ms each. The comment above the function says cat-file --batch-check, but the implementation is per-path ls-tree.

Amplifier: ingest scope is all refs + stash + every reflog-only commit (rev-list --branches --remotes --tags --reflog, ingest.rs:487,512). A worktree's reflog enumeration picked up 2,382 commits vs 853 on HEAD (commits=2224 in the log), ~1.9× the work.

Ask

1. Batch phase B: replace per-path ls-tree spawns with one batched cat-file --batch-check (or one ls-tree per commit listing all paths). Expected: phase B drops from ~70 s to seconds; cold --history build under a minute on this repo.
2. Per-phase timing logs: report/log phases A, B and C separately (tick reports already carry phase; surface a per-phase duration breakdown in the build-index summary so regressions are visible without PIXEL_FACTS_DEBUG_TICKS=1).
3. Scope control: don't index much more than needed — reflog-only commits inflate the work ~1.9× on worktrees and churn-heavy repos. Study what actually pays off in excavate/rescue/history queries and consider bounding reach (e.g. skip reflog-only by default, or an opt-in/out flag), keeping the age window + size budget limits.

Notes

• PIXEL_FACTS_DEBUG_TICKS=1 exists but only prints raw tick reports.
• Reproduction: wipe history.db, run pixel build-index --history ., watch ticks 2–277 sit in phase_b.