> Any ideas on why this happens?

In worst case this can be a bug in wire protocol implementation (very
unlikely), but most likely server just tries to obtain a lock before the
deadlock is reported and does not send any information about the update
outcome to client.

> Is it to do with two-phase commits?

In the same way as it has to do with normal transaction.

In order to confirm the lock issue, change the TPB mapping to isc_tpb_nowait
option, it will return deadlock right after the clash. Alternatively you can
change the deadlock timeout param in Firebird config.

Roman