a raw heading opener with no hard bound renders its title over the whole remainder so the scan is quadratic in a shape the linearity test does not cover

a raw heading opener with no hard bound renders its title over the whole remainder so the scan is quadratic in a shape the linearity test does not cover

Found by the round-10 adversarial ruthless review of build/itd-183. The CLASS is pre-existing -- the parent computes the same title over the same remainder -- but round 10 multiplied its constant, because a title is now rendered from each masking under each of two readings of a removed tag.

When a heading opener has no hard bound, rawHeadingTitleEnds returns len(rest), so renderedTexts and site.Slug run over the whole remainder once per opener. Measured on this round, <p role="heading">x repeated with no close and no h-tag anywhere:

openers   bytes    elapsed
2 000     38 KB    3.3 s
4 000     76 KB    13.5 s

Clean quadratic; the reviewer measured 57 s at 8 000 openers, and at the 4 MiB MaxFileBytes cap this is hours. The parent has the same class roughly 1.5x faster.

TestRawHeadingScanStaysLinearInTheOpenerCount does NOT cover this shape: its openers each close immediately, so 8 000 of them cost 0.1 s against a 10 s bound. The test's name over-claims, and the shape it misses is the expensive one.

Remedies to weigh together: cap the length of text a title may be read from (a heading is not kilobytes long), or give an unbounded opener a bound of the document's own end-of-line rather than end-of-file. Seed material for the exclusion floor's own intent, with the test's name to correct alongside it.

Grounds

• pursued: verifyRedaction finishes in time linear in the document for any run of raw heading openers and refuses nothing the corpus holds; a document up to MaxFileBytes that takes seconds, or a committed markdown file refused for its read budget, would show it wrong