What

InlineParser::buildSingleQuoteMatchCache() decides which single quotes are curly opener/closer pairs and which are apostrophes. It collected potential openers and closers into two lists, then matched each closer against the opener list with a nested scan - O(closers x openers). On quote-heavy text that is quadratic and noticeably slow.

This folds classification and matching into a single forward pass using an opener stack:

if ($prevIsSpace && !$nextIsSpace) {
    $openerStack[] = $i;                      // potential opener
} elseif (!$prevIsSpace && $nextIsSpaceOrPunct && $openerStack) {
    $matched[array_pop($openerStack)] = $i;   // pair with innermost opener
}

The stack top is always the largest-index still-open opener, so popping it reproduces the former "nearest preceding unmatched opener" pairing exactly - now in O(n).

Behavior

Unchanged. The smart-quote semantics, including the official conformance corpus (tests/official/smart.test), render identically; only the cost changes. A new SingleQuoteMatchingTest locks the pairing for matched, lone-opener, nested, and apostrophe-then-pair cases.

This is a perf-only, behavior-preserving change. It deliberately does NOT adopt carve-php's flanking-rule rewrite (which drops opener/closer matching entirely); that diverges from djot.js semantics and would break the conformance corpus. Only the O(n) stack-merge idea is carried over.

Also removes the unused findMatchingSingleQuoteCloser() method, a leftover of the old approach.