@ilevkivskyi you said in #17535 (comment) that you would like to get rid of using types as error contexts. I assume this means this PR is a step in the wrong direction: what exactly should happen instead?

(At the very least I don't think there's something else with the necessary information to improve --pretty's output like this.)

(... reading through the issue, is there a reason why error context can't just include the file? it would be a bit more memory but otherwise seems like a fine solution)

Diff from mypy_primer, showing the effect of this PR on open source code:

sphinx (https://github.com/sphinx-doc/sphinx)
- sphinx/search/__init__.py:198: error: Name "nodes.meta" is not defined  [name-defined]
+ sphinx/search/__init__.py:198:11: error: Name "nodes.meta" is not defined  [name-defined]

pywin32 (https://github.com/mhammond/pywin32)
- win32/Lib/win32timezone.py:913: error: Name "winreg._KeyType" is not defined  [name-defined]
+ win32/Lib/win32timezone.py:913:29: error: Name "winreg._KeyType" is not defined  [name-defined]
- win32/Lib/win32timezone.py:920: error: Name "winreg._KeyType" is not defined  [name-defined]
+ win32/Lib/win32timezone.py:920:19: error: Name "winreg._KeyType" is not defined  [name-defined]

I assume this means this PR is a step in the wrong direction: what exactly should happen instead?

Not necessarily, I think this is a meaningful quick fix. Long term type annotations (essentially what UnboundTypes are) should be part of syntax tree. While actual types should have a base class that doesn't inherit from Context. TBH I didn't thought it through yet, but this is a rough idea.