Memory erase swallows cleanup errors with no retry, so derived data can be left behind #462

Raised by CodeRabbit on #453, but it predates that PR. erase() in sentient/memory/facts.py (~953-955) catches and drops errors from its cleanup steps (vectors, links, summaries, MEMORY.md/USER.md, insight evidence) with no retry. If one step fails (a locked file, a transient DB error), the memory row is gone but some derived data stays, and nothing ever tries again. This affects manual deletes and #441's expiry.

Fix: record failed cleanup steps (for example a small pending_erase table, or a retry list that the hourly memory upkeep drains) and retry them until they succeed. Log them plainly.

Done when: a test that makes one cleanup step fail once shows the derived data is removed on the next upkeep run.