fastparse and native parser assign different line numbers to nested subscripts in multi-line type annotations

[sunny-zuo]

Bug Report

For a multi-line type annotation containing an error in a nested subscript, fastparse and nativeparse assign different line numbers to the resulting UnboundType. fastparse uses the outermost annotation line (self.line fixed at TypeConverter construction); nativeparse uses the actual per-node source line. Downstream, type-arg / arity errors reported against that node show up on different lines depending on which parser was used.

I ran into this as # type: ignore comments in serial mode became unused-ignore when trying parallel mode due to the line number difference.

To Reproduce

producer.py:

def foo() -> dict[
    str,
    dict[int, int, int],
]:
    return {}

mypy.ini:

[mypy]
python_version = 3.12

rm -rf .mypy_cache && mypy --num-workers=0 producer.py                       # fastparse
rm -rf .mypy_cache && mypy --num-workers=0 --native-parser producer.py        # nativeparse, no workers

Expected Behavior

Both parsers should agree on the reported line — either both line 1 (outer annotation) or both line 3 (inner offending node).

Actual Behavior

invocation	reported line
--num-workers=0 (fastparse, no native-parser)	1 (outer dict[)
--num-workers=0 --native-parser	3 (inner dict[int, int, int])

Same error in both: "dict" expects 2 type arguments, but 3 given [type-arg].

Bisection notes:

• The divergence requires the error-bearing node to be a nested sub-expression. Top-level multi-line wrong-arity (x: (\n dict[str, str, str]\n ) = {}) reports line 1 in both parsers.
• Single-line wrong-arity has nothing to diverge on; both parsers agree.

Your Environment

• Mypy version used: 2.1.0 (PyPI). Also reproduces on current master (2.2.0+dev.e53693be05a2).
• Mypy command-line flags: --native-parser selects the inner-line behavior; default (--num-workers=0, no --native-parser`) selects the outer-line behavior.
• Mypy configuration: python_version = 3.12.
• Python version used: 3.12.13

[ilevkivskyi]

This is a known problem. This is probably the biggest deviation between the two parsers. How much trouble does this cause? Potentially we can update fastparse to use actual line in most common cases.

[sunny-zuo]

Apologies for the noise; I did not see an existing issue for this. This is not a major problem, essentially we just had a dozen of # type: ignore comments that needed to be moved when switching over to the native parser. It does mean that we can't switch back to running serial mypy in CI via only a config change, but this is unlikely to be necessary.

Originally I thought this was a divergence between serial and parallel mode which would have been more annoying. Thanks for all of your work on the native parser & parallel mode btw, the speed improvements are fantastic.

[ilevkivskyi]

OK, I see. You can use --native-parser to force using native parser in serial mode (it is a no-op in parallel mode). Then you will be able to use sequential mode on single-core machines without any code changes.