[nicoloboschi]

Fixes #5166.

Problem

Reflect checked the structured memory_ids of done against what it retrieved, but kept the answer text exactly as written. So a full UUID that was made up or garbled was stored in the answer and the mental-model page, and returned 404 when looked up. These IDs often share their first 8 characters with a real ID. Two sources:

• the model writing a wrong ID itself, or
• the length rewrite garbling a citation that done got right (the case raised in the issue comments).

Fix

• ToolResultPresenter records every UUID the model was shown. That covers the raw tool results (so real retrieved IDs, cited by alias or in full, pass) and the opening prompt (question, directives, mission).
• Once the answer is final (after the length rewrite) every UUID outside that set becomes [unverified id]. This runs on the done path (markdown and structured document) and on forced synthesis.
• A UUID quoted inside a memory, like a request id, was seen, so it stays. Unrelated UUIDs are not rejected.

Tests

Each scripts the model and runs the real run_reflect_agent. All three fail on main (the fake ID survives) and pass with the fix:

• done cites a real alias plus a fake ID that shares the real one's prefix. The fake is replaced, and the real ID and a quoted request id stay.
• The length rewrite garbles a citation in a structured document. Both the rendered text and the document blocks are fixed.
• Forced synthesis cites a fake ID.

There is also a presenter unit test.

Not covered

An ID copied from an older version of a page the model read still passes, because the model did see it. Catching that would mean checking every cited ID against the bank.