This more or less looks good to me, though it does seem like a concerning amount of overhead to fix this very specific corner case. Not a big deal though.

though it does seem like a concerning amount of overhead to fix this very specific corner case

Yah, the more specific fix would have been at L498 to check self.ystridx in addition to self.mstridx, but I didn't want to get to a situation where we're checking for ystridx in only one of the many if/else branches, not checking dstridx at all, etc.

The change from 501-511 could be considered orthogonal.

Looks fine. Is the perf really affected? If so then constructing strids once and passing it as an arg (instead of constructing it twice like in the previous version) makes sense.

Aesthetically I like the {"y": self.stridx, ...} better than the (("y", self.ystridx), ...) but that's not a big deal.

I doubt it will really make a major difference in the long run given that the parser is not amazingly performant to start with, but I'd like to minimize the performance impact since this particular bug is for such an incredibly niche problem.

From micro-benchmarks I would estimate I saved a few microseconds by organizing it this way.