I am surprised that this is needed. Up until very recently, inspect was imported eagerly in dataclasses. Isn't the issue with importing dataclasses fixed by the ast change alone?

IIRC the ast change fixes one cycle, but not all of them. If I restore the top-level lazy import inspect in dataclasses, import dataclasses still fails under -X azy_imports=none via: dataclasses -> inspect -> tokenize -> _colorize -> dataclasses

So the inspect move is specifically about avoiding import-time resolution of the lazy inspect import when lazy imports are disabled. The failure is reproducible with a direct import dataclasses, and the stdlib import subprocess test catches it as well.