A commit page highlights only the lines it shows: `highlight` under 15 ms for a 100 KB commit #6187

The commit page now highlights only the lines its hunks draw, taken from the diff itself. It no longer reads any blob and no longer scans whole files. The new tests, the highlight tests and the commit tests pass, and tsc is clean. I haven't measured the 100 KB commit from the issue, and I haven't run the forge's full check.

• How it works: hunkSyntax(diff) in src/web/diff-syntax.ts tokenises each hunk's old side (context and deletions) and new side (context and additions). Tokens are placed at their real line numbers, so the viewer needed no change. A context line that both sides would scan identically is scanned once. src/web/pages/commits.ts uses it in place of diffSyntax.
• Pull page: it still uses whole-file diffSyntax, so its colours are unchanged.
• Tests: test/diff-syntax-hunks.test.ts covers three things.

- One hunk's tokens equal the whole-file tokens at the same line numbers. - Lines outside a hunk are not tokenised. - A commit of 20 files of 800 lines, with ten 3-context hunks per file, highlights in under 15 ms. That test takes the best of five runs, because a single run landed at 10–23 ms.

Colour change: a string or comment opened above a hunk now starts the hunk uncoloured, so its first lines can colour differently from a whole-file pass. I accepted this rather than seeding the tokenizer from the hunk header.

Left out:

• Files that start collapsed (lockfiles and generated files) are still highlighted. Their hunks cost little now, and the viewer's collapse rule is private to diff-view.ts.
• A commit whose hunks are mostly changed lines (about 3,000 lines) still takes roughly 15 ms. The tokenizer costs about 5 µs a line, so the 15 ms target holds only for hunk-sized changes like the test's.

To try it: open any /trevato/forge/commit/<oid> and read highlight;dur= in the Server-Timing header. Run bun test test/diff-syntax-hunks.test.ts.

Closes #6183

The commit page now highlights only the lines its hunks draw, taken from the diff itself. It no longer reads any blob and no longer scans whole files. The new tests, the highlight tests and the commit tests pass, and tsc is clean. I haven't measured the 100 KB commit from the issue, and I haven't run the forge's full check.

• How it works: hunkSyntax(diff) in src/web/diff-syntax.ts tokenises each hunk's old side (context and deletions) and new side (context and additions). Tokens are placed at their real line numbers, so the viewer needed no change. A context line that both sides would scan identically is scanned once. src/web/pages/commits.ts uses it in place of diffSyntax.
• Pull page: it still uses whole-file diffSyntax, so its colours are unchanged.
• Tests: test/diff-syntax-hunks.test.ts covers three things.

- One hunk's tokens equal the whole-file tokens at the same line numbers. - Lines outside a hunk are not tokenised. - A commit of 20 files of 800 lines, with ten 3-context hunks per file, highlights in under 15 ms. That test takes the best of five runs, because a single run landed at 10–23 ms.

Colour change: a string or comment opened above a hunk now starts the hunk uncoloured, so its first lines can colour differently from a whole-file pass. I accepted this rather than seeding the tokenizer from the hunk header.

Left out:

• Files that start collapsed (lockfiles and generated files) are still highlighted. Their hunks cost little now, and the viewer's collapse rule is private to diff-view.ts.
• A commit whose hunks are mostly changed lines (about 3,000 lines) still takes roughly 15 ms. The tokenizer costs about 5 µs a line, so the 15 ms target holds only for hunk-sized changes like the test's.

To try it: open any /trevato/forge/commit/<oid> and read highlight;dur= in the Server-Timing header. Run bun test test/diff-syntax-hunks.test.ts.

I replaced the wall-clock assertion with a count of the work done, pushed it, and could not get a highlight;dur number from the preview.

Test change: test/diff-syntax-hunks.test.ts no longer asserts any timing. For the 20-file commit (800-line files, ten 7-line hunks each) it counts the token rows hunkSyntax produces. It expects exactly 20 × 10 × 7 × 2 = 2,800, where highlighting whole files would give 32,000. Rows come from the hunks only, so the count would fail without the change and cannot flake. The sparse-array and context-sharing checks from before still pass.

Measured note (for the description):

• Warm timing: on a warm run, the same 20-file commit took about 10–16 ms for hunkSyntax.
• Single runs: a single run varied between 10 and 23 ms.
• Tokenizer cost: it costs roughly 5 µs per line.

Preview: I could not get highlight;dur from the preview. /trevato/forge/commit/13cdb24e there, and on production, answers with only app;dur=3–17 and session. It has no highlight phase, so that route is not the commit page in those deployments. The commit timing test, commit-page-timing.test.ts, passes locally. To read the number on a running instance, open a real commit page and look at the Server-Timing header for highlight;dur=. The issue's 71.7 ms figure was from /trevato/forge/commit/13cdb24e on 2026-10-26.