[safa-rk]

is_emergency_number() and connects_to_emergency_number() hand their argument to _extract_possible_number() with no length limit; parse() rejects anything over _MAX_INPUT_STRING_LENGTH before that call, these two don't. _extract_possible_number() strips trailing punctuation with _UNWANTED_END_CHAR_PATTERN.search(), and (?:_|[^#\w])+$ starts a match at every position of a separator run that isn't at the end of the string, consumes the rest of the run and then fails on the $. That makes the call quadratic in the run length: is_emergency_number("1" + "-" * n + "1", "US") takes about 0.28s at n=2000, 4.3s at 8000, 15s at 16000 and 90s at 32000 on CPython 3.9, and these functions are usually fed a raw dialled string. A one-character negative lookbehind on the same class means a match can only start at the beginning of a run, so each run is scanned once: the n=32000 case now takes 13 ms and 1 MB takes about 1s. The leftmost match is always the start of the trailing run, so the match span is unchanged for every input (I compared old and new search() spans on ~636k random and exhaustive short strings). The fix is in the pattern rather than another length cap because PhoneNumberMatcher shares it (its candidates are already bounded, so it isn't affected in practice).

testEmergencyNumberLongSeparatorRun fails on dev (50s, hits the timing assert) and passes in 8 ms with the change. make -C tools/python lint and test pass (309 tests), and the metaclean regeneration leaves no diff.