(Just putting my review summary in the main ticket)

I'm going to echo the previous comment that maybe trying to emulate the existing dict implementation too carefully just adds complexity.

The split-keys implementation shows that there is at least some flexibility to the implementation.  Enough that the keys could map to an array offset, rather than directly to the values?

A simple array of key/value pairs (hashing only to ensure hashability, but ignoring it otherwise) should be faster for the really small dicts you care about, like keyword dicts.  (Admittedly, those timing data are fairly out of date, but I would be surprised if it weren't still true.)