A big issue with this is that although online nodes might be able to detect a long invalid chain, nodes that were offline at the time (or didn't exist yet) have no way of independently verifying that invalid blocks actually existed then. Maybe the invalid blocks were created much later in order to trigger a PoW change earlier than appropriate on some nodes, splitting the network.
You could make a rule that a block is allowed to change the PoW if it presents headers for an invalid chain of length > 50 or something, with the fork point close to the new-PoW block. So once a long invalid chain comes into existence, anyone can create the first new-PoW block containing the invalid chain's headers. This can be more readily verified later on.
Exactly what the new PoW should be is a complicated issue with years of past discussion already...
That is a good point about verifying the historical existence of invalid blocks.
One of the things we can do with the addition of proof of work rather than a complete change is we can compare the hashpower of the SHA256 between different competing chains. Future nodes will be able to assess the health of a the MR POWA fork compared to the 'malicious' chain. If the POWA fork has a substantially lower hashpower it can be assumed the POWA fork failed to achieve economic significance. A proof of work change (POWC) hardfork has no way to relatively compare different chains and determine whether one likely succeeded economically over the other.