It is true that this edge case seems to present an issue, however, the block timestamp will either be valid or not since it is the time since the time embedded in the last valid block with most PoW and not the local time in question.
In the case described, which is not the case but, if may somehow or another be the case otherwise, then some nodes may continue to build upon the block that the other nodes reject. How this would resolve will depend on the percentage of the network split. If 51% or greater of mining power rejects the speculative block then the issue will self-resolve in time as the valid chain is extended further than and having the most PoW. If, on the other hand, 51% or greater of mining power accept the speculative block to build upon then an intervention would be required to pick up the nodes still rejecting or, to discard the invalid chain.
Essentially, if a block complies with the consensus rules (the rules of the reference client) then it is valid.
Others have already stated that it is not wise economically to push the time envelope when mining as it must be decided in advance, so there is an economic disincentive.