Take the log's order from git: torvalds/linux opens in 999 ms instead of 15.7 s

[jonassaa]

Implements #483. Opening torvalds/linux goes from 15.74 s to 999 ms.

Spec: docs/superpowers/specs/2026-09-18-fast-log-walk-design.md
Plan: docs/superpowers/plans/2026-09-18-fast-log-walk.md

Measured

pnpm bench --linux, Apple M4 Pro, idle machine, before and after on the same
day and the same fixture:

	before	after
linux — first screen	15,743 ms	999 ms
linux — first page of history	18,112 ms	222 ms
linux — page ten	15,842 ms	337 ms
deep — first screen	253 ms	59 ms
wide — first screen	5.42 s	5.41 s (it is status, untouched)
refs — first screen	219 ms	224 ms (it is ref enumeration, untouched)

Both acceptance criteria met: the kernel's first screen is under 1.5 s and is
now bounded by status (1.03 s) rather than by history (8.7 ms), and the first
page is under 500 ms.

What changed

Stage 1 — paint before history. refreshAll awaited eleven reads behind one
Promise.all and wrote once, so status, branches, tags and HEAD waited fourteen
extra seconds for the log. The page is still started beside the others (the
eleven-way concurrency open_screen measures is unchanged) but lands in its own
write, and refreshAll still resolves only once history has arrived.

Stage 2 — the order comes from git. build_walk_order now calls
git rev-list --date-order (git/log_walk.rs). Only oids cross over — commit
metadata still comes from libgit2 — so log_cache, FrontierBuilder, cursors
and ref decorations are untouched. Any failure (git missing, non-zero exit,
unparseable output) falls back to today's libgit2 revwalk: a slow page, never a
failed one. git/commit_graph.rs keeps a --split commit-graph warm, written
in the background after open, honouring core.commitGraph.

The correction worth reviewing carefully

#473 and #476 both propose --topo-order. It is the wrong ordering.
Measured byte-for-byte on the kernel's first 2,000 oids:

libgit2 sorting	vs --date-order	vs --topo-order
TIME | TOPOLOGICAL (what log_page uses)	IDENTICAL	differ at line 6; 1,627/2,000 shared

--topo-order does not reorder the same commits — it returns different ones, so
it would have silently changed which commits the first page shows.
tests/log_walk_ordering.rs pins it, with a fixture whose branches interleave on
purpose; the violation was planted and confirmed red, then reverted. The
benchmark's baselines are corrected too.

Two things found on the way

• Oid::from_str accepts abbreviated hex and zero-pads it, so abc123
  parses into a valid-looking oid naming nothing. Truncated rev-list output
  would have become a plausible order with one wrong entry instead of an obvious
  failure. The parser requires full-length hex; a test records why.
• The fixtures now carry a commit-graph, written by bench-fixtures.mjs,
  because the app writes one on open and a fixture without one measures a state
  the app does not leave a repository in. The git baselines get the same
  file, which makes several ratios look worse than when we withheld it. That is
  the correct comparison.

Deliberately not here

• The Settings switch the spec promised. Settings are localStorage; the
  write is a backend decision, so a toggle needs a new Tauri command to write git
  config — a new user-facing surface in a change that already touches the hottest
  read in the app. The opt-out that matters exists and is tested:
  core.commitGraph. The spec records this.
• file_history — 16.9 s on the kernel and now the slowest operation there.
  Same root cause, its own cap, cancellation and cursor semantics (file_history walks all of history on a file with fewer than 500 commits, holding the exclusive lock #474, Cap the file-history walk, share the lock, and let it be cancelled #478).
  log_filtered_page keeps a libgit2 walk on its cache-miss path for the same
  reason. Both earn their own issue.

Verification

• cargo test — 1399 passed, 0 failed (102 binaries)
• pnpm test — 4184 passed (408 files)
• pnpm tsc --noEmit — clean
• pnpm bench --linux on an idle machine; all three published artifacts
  regenerated together and test/benchmark.test.ts re-renders them
• e2e in Docker, snapshot rebuilt — 8 specs passed: smoke,
  history-diff, open-persisted-screen, history-ops, commit,
  status-stage, file-history, repo-tabs

🤖 Generated with Claude Code