do you really need to implement this here in this base class?

if it's just for documenting in the code that some sub-classes will override this, you can just raise NotImplementedError, as you do above

if this should really raise another error, than the error message here is incorrect, as it's not the entire filesystem that is read-only, but only most dirs in it

I know you asked me about this, but now that i rethink of it, permissions on symlinks usually don't matter, because they're never used for anything at all.
Did you have a reason to change this? If not, please revert this change.

don't you need to also raise a FUSEError here, in case of exception? or am I missing the reason why it's not needed?

s/invalidate/purge/

(as "cache invalidation" means a different thing)

This is a very surface-thing test, you're only testing that the directory entry is gone, not that the corresponding entry has actually been removed from the cache (which would be the actual underlying behavior).
Can you add that part, or is it too hidden/not reachable from the outside?

Unrelated question: is the FUSE cache emptied for sure before starting this test? (e.g., by test setup that creates a new instance) Because if it isn't you assertion here might fail if, for reasons unrelated to this test, there are other entries in the cache.