In a memory-off turn, the false-claim correction reveals whether a stored memory backs a claim #551

Found while testing #548. #531 corrects a claimed memory save, but it first checks the stored memories: if one backs the claim, the reply is treated as being about an earlier save and left alone.

In a memory-off turn (/v1 with openai_api.allow_memory off, or incognito with memory off) that check acts as a yes/no oracle. With a scripted model, "your dentist is Kulkarni" got no correction and "Mehta" did, so the app can learn what's stored by watching which claims get corrected.

Fix: when memory_allowed is false for the turn, skip the stored-memory check, so the correction depends only on what happened in the turn.

Done when: a test shows both claims get the same treatment in a memory-off turn, and memory-on turns behave as today.