fix(retain): globalise memory_links lock order on the insert path (#3396)

[nicoloboschi]

Fixes #3396.

The deadlock #2570 targeted still fired a few times a day (per @kubaodias in #3387): DeadlockDetectedError from link_utils during temporal-link insertion, absorbed by the worker retry. Reading both write paths, the insert-side lock ordering was only partial in two ways.

What was wrong

1. The insert sort key covered only half the unique key. _bulk_insert_links sorted on (from_unit_id, to_unit_id) — two of the four columns in the unique index (from, to, link_type, COALESCE(entity_id, nil)). A temporal and a semantic edge on the same (from, to) pair compared equal; Python's stable sort then left them in input order, which isn't stable across transactions, so two concurrent inserts could take the two index entries in opposite orders.

2. Insert and delete used different total orders. chunk_storage.delete_chunks_by_ids orders by (LEAST(from, to), GREATEST(from, to), link_type, COALESCE(entity_id, nil)) — direction-normalised so (A,B) and (B,A) sort adjacent. The insert sorted on the raw (from, to), placing them far apart. Both orders are internally total but different, and a consistent lock order requires every writer to share one.

The fix

Both call sites now sort on one canonical key — the full, direction-normalised unique key. Extracted _lock_order_key and pointed the insert sort at it; the delete side already uses exactly this order and is unchanged.

(LEAST(from, to), GREATEST(from, to), link_type, COALESCE(entity_id, nil))

UUID string ordering matches Postgres uuid byte ordering because the ids are canonical lowercase-hex (the same assumption the old str() sort already relied on).

Tests

TestLockOrderKey in tests/test_link_utils.py — 4 pure unit tests: direction normalisation, link_type disambiguation (mechanism 1), None → nil-UUID mapping, and that a mixed batch reproduces the delete's ORDER BY. Full file: 31/31 pass; ruff + ty clean.

Note

As the issue says, this is derived from reading the two paths, not a reproduction — it closes both ordering gaps rather than betting on which one fires in production. Severity is low (a few/day, fully absorbed by the worker retry); this makes the #2570 ordering guarantee actually global.

[kubaodias]

Heads-up: this crossed with new evidence on #3396 — server-log DETAIL blocks plus a local repro show the firing cycle waits on memory_units tuples (delete-side X locks from earlier statements in the retain txn vs the locked CTE's FOR KEY SHARE), not on memory_links index entries, so the sort-key change here won't reduce the observed deadlocks. Details in #3396 (comment). The change itself looks harmless and the ordering unification is reasonable hygiene — but it may be worth softening "Fixes #3396" to "Refs #3396" so the issue stays open for the cross-statement cycle.