[BUG] Model passes an unverified generalisation from memory to subagents as fact

The model instructed review subagents that a protocol-wide HTTP status rule was fact, taken from its memory notes. For one service the botocore model contradicts it (the not-found error is modelled with HTTP 404). A subagent had to correct the premise. The session's own instructions say memory contents are claims that must be verified before being asserted.

Expected: premises handed to subagents are verified or labelled as unverified.