Lift the 128 KB Vale file cap: 3.21.0 makes single-block lint time linear

[thecodedrift]

What

Lift the 128 KB oversized-file guard, VALE_MAX_FILE_BYTES in packages/cli/src/rules/vale/run.ts:120 (introduced by #323), now that the CLI vendors Vale 3.21.0 (#336).

Why it existed

#325 measured, on Vale 3.20.0, that lint time was superlinear in the size of a single Markdown block: a 3 MB one-block file took ~80 s while the same sentences with blank lines between them took ~4 s, and a 128 KB single block cost ~194 ms, 512 KB ~2 s. Combined with Vale emitting nothing until every file is done, one oversized block under a time limit cost the whole run its findings. File size was the proxy we could afford to measure before spawning, so #323 capped it at 128 KB.

What changed

Vale 3.21.0 ships two perf commits (ed711769 index the walker's context; 252d0963 index rune positions and resume column counts per block). The Vale upgrade in #339 re-measured our own single-block fixtures on the vendored binary (see the tables in packages/cli/test/vale-run.test.ts):

single block	3.20.0	3.21.0
152 KB	~1,020 ms	~45 ms
323 KB	~4,400 ms	~67 ms
6.7 MB	—	~1,000 ms
19 MB	—	~2,800 ms

Cost is now linear at ~2.8 µs per repetition, block structure no longer matters, and the 128 KB line no longer separates a cheap file from an expensive one: 128 KB of one block costs tens of milliseconds.

What to do

1. Re-run Track upstream: Vale lint time is superlinear in single-block size, and no output is emitted until the run finishes #325's reproduction (its 3 MB single-block huge.md) against the vendored 3.21.0 to confirm on that exact shape, since the issue says the repro is ours to re-run on each vendored Vale. Record the number here.
2. Remove VALE_MAX_FILE_BYTES and the per-run override that wraps it (run.ts:546), the skip/notice path in formats.ts (:317, :322, :560) that reports a file as oversized, and their tests. If we would rather keep a guard against a genuinely pathological input, raise it to something a linear cost justifies and say in the docblock what it now protects against; the current docblock's justification is the superlinear measurement, which is gone.
3. Update the update.md ledger only if a user could observe a file that was previously skipped now being linted — they can, so one ledger line: files over 128 KB are linted again.

Still true

Vale still writes nothing until the run finishes, so a run that hits an external time limit still loses every finding. That half of #325 is not fixed upstream; it is just far harder to reach when cost is linear. Keep the timeout test's headroom reasoning in vale-run.test.ts as-is.

Refs #325
Refs #323

[thecodedrift]

Duplicate of #351 (same fix, that one carries the decision to land it in the next patch). The measurements and file locations here stay as the reference.