fix(rewrite): keep the minus sign on a line-initial negative number (#821)

[Vaishnavi1709]

Summary

cleanRewriteLine stripped the minus off a line-initial negative number, so -5% churn reached the résumé — and the exported ATS PDF — as 5% churn. A reduction read as a gain, and nothing downstream looked wrong: the line stayed grammatical and the figure stayed plausible.

LIST_MARKER_PATTERN folded - and • into one [•\-]\s* branch accepting zero trailing whitespace. Inside the fixed-point loop from #806, pass 1 removed the - marker and pass 2 read the uncovered - as a bullet glyph and removed it too; the bare -5% shape needed only one pass and predates the loop. This is the dash-side twin of the decimal bug #806 fixed on the numbered branch — the loop's own docblock warns that "any strip added to this loop must be checked the same way", and the -/• branch was carried over without that check.

The fix splits the two glyphs apart and withholds the zero-space branch from - when a number follows:

-const LIST_MARKER_PATTERN = /^(?:\d+[.)]\s+|[•\-]\s*|\*\s+)/;
+const LIST_MARKER_PATTERN = /^(?:\d+[.)]\s+|•\s*|-(?:\s+|(?!\.?\d))|\*\s+)/;

• keeps \s* because it is only ever a glyph, never part of the text it precedes. - is the one marker that doubles as ordinary content.

Closes #821

Deviation from the fix proposed on the issue

Worth flagging explicitly, since the issue body specifies a different regex.

The suggested -(?:\s+|(?=[A-Za-z])) regresses an existing test: it stops stripping the tight -**Led** the migration of billing. case pinned at post-process.test.ts:194 ("tight dash, no space"), because * is not a letter, so the lookahead fails and the marker survives. Caught by simulating the fixed-point loop over every existing marker case before writing code.

A negative lookahead on the number is the property actually wanted, and (?!\.?\d) additionally covers -.5% as well as -5%.

Review focus

• src/lib/webllm/post-process.ts:127 — the \s+ branch is deliberately left unguarded, on the reasoning that - 5% is a marker plus a positive number and a genuine negative is written - -5%. Is there a shape where a model emits - followed by an intended negative with no second sign?
• src/lib/webllm/post-process.ts:127 — (?!\.?\d) is a lookahead, so a tight - before a non-number still strips as before. Does that hold for every non-letter opener the rewrite path can emit (-**Led**, -"quoted", -(paren))?

Test plan

• npm run typecheck clean
• npm run lint clean
• npm run verify green (also re-run by the pre-push hook)
• Full suite green — 383 files / 6457 tests passed, 8 skipped. Run explicitly rather than relying on verify's scoped test:changed selection.
• New cases falsified against the pre-fix pattern — reverting only the regex turns all 7 new assertions red, the other 51 in the file stay green.
• All four repro lines from the issue fixed; -Shipped X, -**Led**, • 4.2%, - 3.5x, 1. Foo and *Foo.* unchanged.

Six negative-number rows join the decimal-preservation it.each table — the table that already asserts this property for the numbered branch — plus a sign-preservation property test across all five marker shapes.

No fixture binaries touched, so the fixture-PII preflight does not apply.

[cloudflare-workers-and-pages]

Deploying offlinecv with    Cloudflare Pages

Latest commit:	f152fae
Status:	✅  Deploy successful!
Preview URL:	https://5ca556be.offlinecv.pages.dev
Branch Preview URL:	https://fix-821-negative-number-minu.offlinecv.pages.dev

View logs