On Fri Jun 27 15:32:51 2025 +0000, Huw Davies wrote:
I spent a bit of time looking at the tests in the first commit. The error term passed to some of the calls to `match_off_by_n()` is very large, which means they're hiding the fact that the proposed fixes aren't correct. We should be able to get this to match to within a pixel or two. I don't have a suggestion on what the correct fix would be, but it's clear that this isn't right as it stands. It might be easier (at least for development) to start with simple rotations and once those are understood, add back the shear.
Thanks for looking at it. As explained in the comment in the tests Windows produces noticeable diff for angles 45.0 and 135.0, so very large margin is to make tests pass under Windows. Should I remove the tests for angles 45 and 135, or do something else?