Exactly. In the case of the most recent fork, it was a completely unexpected implementation bug that meant that the underlying database ran out of locks, different to the more recent builds that didn't. That was caused just by differences in underlying databases, let alone a whole new verification system as we see here.

Obviously this is not a desirable property of the system. If it's going to be robust against malicious attacks, it must be robust against well-meaning reimplementations.

Malicious attacks are not the same as consensus failure.

Consensus requires agreement, malicious attacks are orthogonal, unless you're hypothesizing a malicious party introducing an intentionally incompatible node and a significant fraction of users switching to it.

What does it even mean for Bitcoin to be robust against malicious attacks? Bitcoin has no security definition, just a vague notion of what its users should not be able to do.