Certified Panic Mode: Repair-Invariant Error Recovery for Maximal-Munch Lexing

Abstract

Classical panic-mode recovery skips a failed scan forward by convention: it promises progress, and membership in a designated set where one is found, but no repair-universal boundary guarantee for the position it lands on. This paper chooses the resume position by theorem: a position is a sound recovery point when every prefix repair whose scan commits through the certificate's returned evidence places a token boundary there, complete repairs the special case, by a committed-prefix lemma over certificates posted for a different purpose (arXiv:2608.03473, arXiv:2608.09761). The quantifier is strictly the stronger one, and a dichotomy locates the difference exactly: the inclusion is strict precisely when no repair completes while some scan still commits through the evidence, a two-token witness realizing the case. The evidence-returning form pairs each answer with the interval it rests on, which is what makes the guarantee deployable: a caller who knows what the scanner cannot check decides whether the certified boundary transfers to the clean input it intended. The search is one forward walk in evidence order and provably terminates; the guarantee is per-automaton, relative to the active mode. The procedure ships in the munch lexing library, and a deterministic corruption study measures its recovery quality beside the classical skip-one and delimiter conventions. Every certified answer whose evidence survives the damage passes its executable landing check on every recovery move, 40,885 of 40,885; among first answers, the 5,595 resting on evidence the damage reached land 1,589 times, reported and never asserted. On 13,522 of the 16,808 trials the shipped repair routine labels beyond repair, the resumed suffix tokenizes whole and the two quantifiers provably coincide, so both are empty there if those negative labels are exact; 3,286 stay undetermined.