docs(evidence): correct S5 — the semver-parse survivor is an equivalent mutant, not a weak witness

[IanFrelinger]

Corrects S5. The claim that semver-parse's mutation survivor was a weak witness was wrong, and running it down produced the more interesting result.

What the survivor actually is

Survivor ids are mutate-int-literal-{line} (AstMutationCatalog), so mutate-int-literal-41 is line 41 of the candidate:

if(version.Length>0)// catalog rewrites 0 => 2, giving: version.Length > 2

The original and the mutant differ only for inputs of length 1 or 2. No such input can be a valid semver: isValid is set only when corePart.Split('.') yields three parts that each pass IsValidNumericPart, and the sole length-≤2 string containing two dots is "..", whose parts are empty and rejected. On every input that distinguishes them, both return isValid=false, 0, 0, 0, "".

The guard is redundant. The mutant is semantically equivalent, so no witness case can ever kill it — and since CertificationGate rejects on EscapeRate > 0, this candidate could not have certified regardless of what was added to the witness or how many repair attempts it got.

BrickMutationEngine already discounts three classes of unkillable mutant — source-identical, non-compiling, and analyzer-dead. Semantic equivalence is a fourth that it does not model, and is undecidable in general.

This is the first time a dogfood campaign has caught the trust machinery being wrong rather than the proposer, which makes it the substantive finding of S5.

Knock-on correction to the repair claim

S5 said repair re-emission held at both sizes. It does not follow at 27B: 2 of that model's 4 repair attempts were semver-parse, where no edit could have helped — so identical re-emission there was the correct response, not a failure to read the feedback. That leaves text-slug as the only valid test at 27B, where it did edit once in two attempts. One objective is not a measurement, so this is now recorded as untested rather than answered.

The 7B evidence is unaffected and stands: qwen2.5-coder:7b re-emitted byte-identical source on 6/6 repair attempts across three objectives that all had genuine, fixable defects.

Docs only — no behaviour change. The gate fix, if any, is a separate decision: EscapeRate > 0 as a hard reject is unsound in the presence of equivalent mutants, but every remedy has real trade-offs.

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.