fix: correct the comparison relation sentence and remove NUL bytes from source

[renrenmimi]

Two correctness fixes to the comparison feature. No UI redesign, no summaries.

Problem 1: the accessible sentence named the wrong subject

The visually hidden summary dropped the relation word into a fixed
X is <word> of Y frame. That frame was backwards, and ungrammatical for two of
the four relations.

Before

99e7880 is ahead of 49b7926: 2 commits ahead, 0 behind, 12 files changed, 714 lines added and 170 removed.
49b7926 is behind of 99e7880: 0 commits ahead, 2 behind, 12 files changed, 170 lines added and 714 removed.
49b7926 is identical of 49b7926: 0 commits ahead, 0 behind, 0 files changed, 0 lines added and 0 removed.
aaa1111 is diverged of bbb2222: 3 commits ahead, 5 behind, 20 files changed, 300 lines added and 200 removed.

In base...head semantics aheadBy is how far head is ahead of base. The
first line therefore named the wrong commit: 49b7926 is the one that is two
commits ahead, not 99e7880. The trailing clause repeated the same numbers with
the same ambiguity, and is behind of / is identical of / is diverged of are
not English.

After

49b7926 is 2 commits ahead of 99e7880. 12 files changed, 714 lines added and 170 removed.
99e7880 is 2 commits behind 49b7926. 12 files changed, 170 lines added and 714 removed.
49b7926 and 49b7926 identify the same commit. No file differs between them.
bbb2222 is 3 commits ahead of and 5 commits behind aaa1111. 20 files changed, 300 lines added and 200 removed.

Each relation is now a complete sentence of its own, so the grammar cannot change
meaning between statuses. Singular and plural are handled (1 commit ahead), and
thousands are separated to match the numbers on screen.

Sighted readers get the direction from the arrow between the two endpoint cards.
That arrow was the only thing saying which way the numbers ran, which is exactly
why the sentence has to say it in words.

The wording is a pure function in src/lib/compare/describe.ts.

Tests

Written against real comparisons of a known history, so the endpoint order is
checked rather than assumed:

• head is the subject and base the object, asserted by string position, not by
  the presence of the words;
• ahead, with the exact sentence;
• behind after swapping the ends, with the exact sentence, and asserting the
  old behind of shape is gone;
• identical, including the changes clause;
• diverged, with both distances;
• a guard that the old reversed template can no longer be produced;
• singular nouns and thousands separators.

E2E reads both shas off the endpoint cards and asserts the sentence names them in
the right order, before and after pressing swap.

Problem 2: raw NUL bytes made source files binary

ComparePicker.tsx used a raw NUL as a sentinel for "nothing selected", as the
fallback of selected ?? ... inside a startsWith call.

It behaved correctly, since no sha starts with a NUL. But it put two raw NUL
bytes into a TSX file:

$ file src/components/ComparePicker.tsx
src/components/ComparePicker.tsx: data

$ git diff --numstat
-	-	src/components/ComparePicker.tsx

Git classified the file as binary and stopped producing readable diffs for it. A
change nobody can read in a diff is a change nobody can review.

The intent is now stated outright:

const isSelected = (sha: string): boolean =>
  selected !== null && sha.startsWith(selected);

After

$ file src/components/ComparePicker.tsx
src/components/ComparePicker.tsx: Java source, Unicode text, UTF-8 text

$ git diff --numstat        # after a one-line probe edit
2	0	src/components/ComparePicker.tsx

The guard found two more

A source-integrity test was added, and it immediately found two files carrying
raw NUL bytes from earlier work that I did not know about:

File	Why a NUL was there
e2e/api.spec.ts	hostile test input for path validation
src/lib/stats/growth.ts	separator for a composite area + bucket map key

Both intents are sound, so both keep their NUL, written as a unicode escape
sequence instead of embedded as a byte. The runtime strings are identical, so
behaviour is unchanged; the files are now text. All three report as text to
file and to git grep -I.

The guard fails if any tracked source file gains a NUL byte, fails to decode as
UTF-8, or becomes binary to Git. It was verified in both directions: injecting a
NUL makes it fail, removing it makes it pass. It also caught itself once, because
the replacement-character check contained a literal U+FFFD, which is the third
commit.

Picker tests

23 component tests covering the acceptance criteria and the surrounding
behaviour that had to stay intact: a null selection marking nothing, full and
abbreviated sha selection, tag selection, tags outside the loaded range not being
offered, filtering by subject, sha and tag name, Escape, outside click, toggle,
disabled, focus moving to the filter field on open, and the side labelling.

E2E keeps its 360px picker-open overflow check.

Verification

Check	Result
npm run typecheck	clean
npm run lint	clean
npm test	533 passed (was 496)
npm run build	clean
npm run test:e2e	159 passed (was 157)
privacy audit	3/3 pass
token canary	no canary in any response, header or browser chunk

Preserved

Built-in demo unchanged: dataSource: builtin, htmlUrl: null, exactly 16
commits, rateLimit: null and zero GitHub requests. Unknown demo/* still
refused locally with 404 and null quota fields. Badge still reads
0 GitHub requests. src/lib/github/client.ts untouched, so the server-only
token architecture is byte-identical. Caching and rate-limit behaviour unchanged.
No generated summaries anywhere.

No old PR description was modified, no history rewritten, no force-push.

Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
repo-time-machine	Ready	Preview	Aug 31, 2026 11:34pm UTC

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a28d3eb2d4

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Include tracked .mts sources in the integrity guard

The extension allowlist omits .mts, so the tracked TypeScript source vitest.config.mts is never checked. If that configuration file acquires a raw NUL or invalid UTF-8, this new guard still passes even though Git may classify the file as binary—the exact regression the test is intended to prevent. Include .mts in the filter.

Useful? React with 👍 / 👎.