The elements in the block filter are de-duplicated on insertion into a set, however the hashes are not. This means a zero is encoded in the filter for each collision. The client never cares about these so they should not be needed.
Is this actually a bug or am I missing something?
For reference, the same issue is present in the Rust library.
It appears to be a bug (or at least an inefficiency), but I think the practical implications are pretty minimal, as collisions should only occur about 1/784931 of the time on average for the basic filter.
It does appear that the BIP calls for a "set" of the hashes, which implies no duplicates, but it's not explicit.
This is not a Bitcoin Core problem and should be addressed on the mailing list to update the BIP.