Discussed with @BramGruneir and team. The current semantics in CRDB 2.1 are incorrect -- the FK work must be deferred to no earlier than the end of the current statement.

In the first iteration of this we are reusing the same FK code; and instead delay the processing until the end of the statement's execution by means of callbacks.

Also the set of modified rows must be collected using a "disk row writer" that spills to disk to limit RAM usage.