[nicoloboschi]

Fixes #5804, #5806, #5830. Refs #5805 (made diagnosable, not fixed).

#5830: creating a mental model on a new bank returned a 500 on a deadlock (real bug)

The CI deadlock detail: one process waits for ShareLock on memory_units, the other for ShareLock on a virtual transaction. Creating a bank builds its vector indexes with a plain CREATE INDEX inside the caller's transaction. Another bank's CREATE/DROP INDEX CONCURRENTLY waits on that transaction, and the transaction waits on it back. create_bank_if_missing already retries this. The three writes that create the bank on their own connection did not: create_mental_model, create_knowledge_folder and create_knowledge_page. All three now run their transaction under retry_with_backoff. It rolled back, so a redo is safe. Indexing the page still happens after the commit.

Regression test test_write_that_creates_its_bank_retries_transient_deadlock (one case per path) injects one DeadlockDetectedError. All 3 cases fail without the fix and pass with it.

#5804: event-loop test failed on one machine stall

It asserted on the single longest gap, so one GC pause or CPU-steal stall on CI could fail it. It now counts the ticker's turns during the batch (>= 3). Checked: with the resolver's asyncio.sleep(0) removed, the test fails with 0 turns.

#5806: re-ingest + delete of linked documents deadlock

Neither operation runs index DDL. But planning any memory_units statement takes AccessShare on every bank's vector index. That means another xdist worker's CONCURRENTLY drop can deadlock with either side (the same pattern as #5752). The test now reruns a round, up to 3 times, only when the deadlock detail names a virtual transaction. A real row-lock cycle between the two operations still fails, and the failure now prints the deadlock detail. The CI log had none, so I can't prove which kind it was.

#5805: not fixed, made diagnosable

From the CI log: DELETE FROM documents in delete_bank hit the 60s command timeout (TimeoutError, empty message). During that same minute almost every other xdist worker stalled too. So it looks like a DB-wide lock queue, not something in this test. A likely chain is bank-create's plain CREATE INDEX ShareLock queued behind a CONCURRENTLY index build, with every writer queued behind it. Not proven. The error now carries the type (TimeoutError: ...) and the cause, so the next failure tells us more. The lasting fix is likely to stop building a new bank's indexes with a plain CREATE INDEX inside a transaction. That is bigger and left for a follow-up.

Also closed

• Flaky: test_later_message_wins_across_overlapping_sessions[update-at-last-message] #5773: not a flake on the CI model (see fix: flaky tests #5752, #5764 (deadlock retry on re-retain, delta prompt) #5784).
• Flaky: test_import_from_a_file_in_small_batches imports fewer observations/documents than exported (and can deadlock) #5781: no longer reproduces on main (3/3 pass, was 3/3 fail). Likely fixed by fix(transfer): export and import a bank's growing tables a page at a time (#5737) #5792.