fix: perf quadratic ^block-id transform on long lines

[zeyutang]

Problem

transformMarkdown runs the ^block-id replace on every normal line, ungated:

line = line.replace(/(.*?)\s+\^([a-zA-Z0-9_-]+)$/, ...);

The lazy (.*?) head under the $ anchor backtracks quadratically on every non-matching line, which could mean every sentence of a real document in the worst case. Cost grows with the sum of squared line lengths, so large documents with long lines (e.g. one sentence per line, without manual line wrap) reach multi-second previews: a CPU profile of a 192 KB paper attributes ~93% of warm parseMD self-time to this one regex.

Fix

Gate the replace behind a linear-time suffix test, leaving the replace untouched:

if (/\s\^[a-zA-Z0-9_-]+$/.test(line)) { ... }

The guard matches iff the replace regex matches (both require the line to end with whitespace, then ^, then one or more id characters), and when both match the original replace still runs, so rendered HTML is byte-identical. A greedy rewrite (/^(.*)\s+\^([a-zA-Z0-9_-]+)$/) would also be fast but shifts part of a multi-space run into group 1, changing output whitespace.

Timing (same machine, Node v26.4.0)	before	after
warm parseMD, 192 KB one-sentence-per-line file	~2.1 s	~55 ms
smoke test, 2000 non-matching lines x ~2128 chars	~5.6 s	~10 ms

Tests

• Fixture pair block-id.md / block-id.expect.md covering the syntax battery (transform positives, mid-line and no-space negatives, $$ math and code-fence protection, indented code), pinning byte-identical output.
• Unit tests for cases whose bytes editors and git rewrite in fixture files: trailing space after the id, CRLF, multiple carets (a ^x ^y), whitespace-only prefix.
• Long-line smoke test with no asserted time bound (e.g., ~10 ms guarded vs ~5.6 s unguarded).