set_precision() can cause negative buffer on polygon to fail

Edited to include OS, shapely versions, etc.

Me again, with another unexpected set_precisionside-effect (I'm working on some low-level geometry stuff and have run into a bunch of these as reported in issues #1947, #1950, and #1952.)

Expected behavior and actual behavior

Setting precision on a polygon should not affect the correct working of a negative buffer, but in some instances (example below) it does.

This one is particularly strange and unpredictable. I tried to make an MRE with a nice simple square (0, 0) to (100, 100), but that didn't work, so I am presenting it here with a polygon that fails on my system. It's a cairo polygon for anyone interested.

Steps to reproduce the problem

import shapely.ops
import shapely.geometry

p = shapely.geometry.Polygon(
  [(-250.0, 0.0), (-197.168784, 197.168784),
   (0.0, 250.0), (52.831216, 52.831216),
   (-52.831216, -52.831216), (-250.0, 0.0)])
p = shapely.set_precision(p, grid_size = 1e-6)
p = p.buffer(-1e-5)
print(p)

which yields an empty polygon:

POLYGON EMPTY

The buffer does not fail at all sizes, and it's not predictable which sizes it will fail at: 1e-6 works fine, 1e-7 fails) For the most part very small buffers seem to fail, while larger ones succeed. I assume this is some strange floating point effect, but am unsure. I found it quite by chance and as I have noted, it doesn't fail on a simple square polygon.

Importantly the buffer doesn't fail at any size I have tested if I leave out the set_precision step.

I encountered this problem when using a negative buffer in data cleaning (where I had the buffer style set to minimise the buffer segments, and thinking it might relate to quantisation issues with a coarse buffer style and small numbers changed to 'default' buffering by not setting any of quad_segs, or join_style, etc., but as above the issue persists.

Operating system

Mac Sonoma 14.2.1

Shapely version and provenance

pip installed 2.0.2

Start by running the provided set_precision() and negative buffer reproduction with the reported Shapely 2.0.2 setup, then compare buffer sizes with and without precision setting. Trace the set_precision() and buffer entry points to identify the interaction, and add a regression test showing that the negative buffer does not unexpectedly become empty.