docs: add cascade delete design decision to multi-backend plan

DataJoint keeps its custom cascade delete implementation rather than
using native ON DELETE CASCADE for dry-run capability and to handle
complex foreign key graphs with converging/diverging references.

Author: claude

docs/design/multi-backend-plan.md

@@ -145,6 +145,7 @@ Replaced MySQL-specific `NOW()` with standard SQL `CURRENT_TIMESTAMP` in:
 | **datetime (not timestamp)** | Matches Python naming; familiar to scientists; avoids MySQL/PG timestamp confusion |
 | **json → jsonb on PostgreSQL** | Better performance and indexing; key order not guaranteed by JSON spec anyway |
 | **int8 → smallint on PostgreSQL** | PostgreSQL lacks tinyint; smallint is semantically equivalent (small integer) |
+| **Custom cascade delete** | DataJoint implements its own cascade delete (not `ON DELETE CASCADE`) to provide dry-run capability and handle complex foreign key graphs with converging/diverging references. This also ensures consistent behavior across backends. |
 
 ---