ResponseAwareGuard

A MatchGuard that also reads the candidate's stored response, not just the two prompts.

Every other guard compares a query with a cached prompt, because that is all the match path used to offer. That leaves one shape of near miss structurally invisible: two prompts that read as honest paraphrases, where the thing separating the answers lives in the answer. "What is the capital gains tax rate when I sell a second home" and "…a primary residence" clear every guard in MatchGuards.standard — no number differs, no unit, no negation, nothing either prompt says is evidence. The cached answer opens "Gain on a second home is taxable in full", and serving it to the other question is a confident, wrong answer about somebody's tax bill.

Only the candidate's response exists at this point. The query's answer is precisely what the cache is trying not to pay for, so a response-aware guard cannot compare two answers. It can only ask whether this answer was written for this question.

The prompt-only evaluate abstains, so dropping one of these into a plain List<MatchGuard> — or into a call path that has no entry to hand, such as dev.kmemo.SemanticCache.explain on a store that returned no response — is safe rather than silently wrong. Abstaining is always the right move over guessing; see MatchGuard for why the cost of the two mistakes is not symmetric.

Implementations carry the same contract as any guard: pure, thread-safe, and fast enough to run on every candidate above the threshold. A response can be far longer than a prompt, so "fast" is the binding constraint here — read the response, do not parse it.

Inheritors