Consider switching to Pyrefly for typechecking?

[javabster]

I'd like to propose adding Pyrefly for typechecking - it could either be added alongside mypy or migrated entirely. Happy to open the PR, just checking here first to to see if there's interest. For transparency I work on the Pyrefly team, I came across this project while investigating the performance of mypy vs Pyrefly on real repos and it looked like it could be a good fit :)

I did some initial investigation, running locally mypy 2.3.0 vs Pyrefly 1.1.1:

Checker	Cold run	Warm run
mypy strict	9.2 s	0.28 s (.mypy_cache)
Pyrefly	0.6 s	0.32 s

~15× faster cold, and Pyrefly stays sub-second with no separate cache step required

How to make the switch:

1. Add the dependency to [tool.poetry.group.typing.dependencies]: pyrefly = ">=1.1.1".

2. Add [tool.pyrefly] config mirroring the existing [tool.mypy] config:

   [tool.pyrefly]
   project-includes = ["src", "tests"]
   search-path = ["src", "."]          # mirrors mypy_path=src + explicit_package_bases
   project-excludes = ["**/__pycache__", "tests/fixtures",
                       "tests/masonry/builders/fixtures", "tests/utils/fixtures"]
   replace-imports-with-any = ["fastjsonschema.*", "findpython.*",   # mirrors mypy's
                               "requests_toolbelt.*", "shellingham.*", "xattr.*"]  # ignore_missing_imports

3. There are some differences between the errors pyrefly catches and what mypy does (38 new errors from what I can see running locally, details below), so we could either ignore them (# pyrefly: ignore [code]) for now or see if there are genuine issues to address. If we opt for removing mypy we could also remove ~30 mypy ignores that pyrefly doesn't complain about.

4. Wire into CI in .github/workflows/.tests-matrix.yaml, mirroring the mypy job
   with a run pyrefly check step. Suggest starting non-blocking, then flipping to
   required once you're comfortable.

Error differences - Mypy vs Pyrefly

When I compared pyrefly and mypy output, I found 38 new errors from Pyrefly that aren't flagged by Mypy:

Error code	Count	Reason for difference	Suggested approach
missing-attribute	10	5 = Windows-only stdlib (winreg/windll) checked on non-Windows (pyrefly doesn't narrow sys.platform); 5 = descriptor/mock/stdlib-API patterns	Suppress
unbound-name	7	Local possibly used before assignment; mypy strict doesn't run this check	Suppress - all looked benign, optionally could cleanup
bad-override	7	Pyrefly has stricter override/LSP checks than mypy	Suppress - all errors occurred in tests
bad-assignment	5	Container invariance + descriptor + test-typing mismatches	Suppress
bad-argument-type	3	Looks like real type mismatches (`Term	None→Term`, etc.)
bad-override-mutable-attribute	3	Mutable-attribute variance; 1 in /src, 2 /test	Suppress tests and possibly fix /src
bad-return	1	Return type wider than declared annotation, possibly a real type mismatch	Review/fix
unsupported-operation	1	Item assignment on type without __setitem__ (BaseSpec)	Review/fix
bad-index	1	Indexing non-subscriptable EntryPoints	Suppress - only in test
Total	38		~33 suppress, ~5 fix/review