Allowing small allocations to fail

> The only time a kernel memory commit can fail is if too much memory has already been granted to userspace.

Only if you include memory used for buffering writes to disk as "granted to userspace". And that memory is not something that overcommit or copy-on-write will affect in any way.

Remember that failures like this are pretty rare, if a system has even a 'little' swap (say a couple GB), it will almost certainly slow to an unusable dead crawl from the disk I/O long before the OOM killer kicks in.