Publish the measured performance numbers in the README

[jonassaa]

Part of #257 — the README half. The site half is still deliberately waiting, and docs/dev/performance.md now says why the two differ.

What

A ## Performance section in the README, between "How it compares" and "Features": one row per benchmark fixture, with the first screen, working-tree status, the first page of history and the ratio against git's own work.

Repository	First screen	Status	First page of history	…vs git
torvalds/linux	15.84 s	989 ms	15.95 s	1.7×
deep	253 ms	0.53 ms	249 ms	1.3×
wide	5.42 s	5.42 s	0.25 ms	—
refs	219 ms	0.55 ms	135 ms	16×

Nothing was re-measured. The block is rendered from the committed docs/dev/benchmark.json, which is the same record docs/dev/performance.md's tables come from.

Why it is generated rather than typed

The README's first line calls this app fast, which makes it the single place where nudging a figure would pay off most — the failure #257 exists to prevent. So scripts/bench-report.mjs gained renderReadme() and a marker pair of its own, pnpm bench now writes three files instead of two, and test/benchmark.test.ts re-renders the README block from the record and fails on any disagreement.

Four new guards, each proven by planting the violation it describes and confirming it failed alone:

guard	planted violation	caught
re-renders the committed README block	changed 253 ms → 153 ms by hand	✅
a real timing in every README cell	swapped a column to an op two fixtures lack	✅
no hand-written figures in the prose	wrote "…is 253 ms" above the markers	✅
README.md in the js path filter	deleted the filter entry	✅

The third is the one worth calling out: the prose around the block is hand-written, and a measured figure copied into it is exactly what stops moving on the next run. 50,000-commit does not trip it; 253 ms does.

It leads with the bad case

performance.md said publishing 15.8 s for torvalds/linux as a selling point was the wrong response to it. That reasoning is about the marketing site and still holds there — but the README already carries "Where we are behind" and a "Status" list of known gaps, so the honest number belongs on it. The paragraph under the table names it outright, and it is interpolated from the record rather than typed, so fixing the log walk also unwrites the sentence.

docs/dev/performance.md is updated to distinguish the two audiences instead of saying "not published yet".

Verification

• pnpm tsc --noEmit — clean
• pnpm test — 4182 tests in 407 files, all passing, run after the last edit
• Four planted violations, each failing its own test and only its own

No src/ or src-tauri/ change, so e2e is untouched.

🤖 Generated with Claude Code