False positive involving `re.sub` and `+` operator introduced in mypy 0.960

[DMRobertson]

Bug Report

Consider the following snippet. (This is a minified example of the problem detected in matrix-org/synapse#12996.)

import re

"#" + re.sub(
    "PATTERN",
    lambda match: match.group(0) + "ANOTHER_STRING",
    "HAYSTACK",
)

Mypy 0.950 is happy with this, but mypy 0.960 is not.

main.py:5: error: Unsupported operand types for + ("bytes" and "str")

AFAICS this is a legitimate usage of re.sub, consistent with the stubs defined in typeshed.

Additional

Strangely, removing either of the addition operators and the corresponding string literal causes mypy 0.960 to accept the snippet. That is, applying this change

- "#" + re.sub(
+ re.sub(

is accepted, and applying this change

-    lambda match: match.group(0) + "ANOTHER_STRING",
+    lambda match: match.group(0),

to the vanilla snippet above is also accepted.

I'm not sure how to get mypy to reveal the type of an expression in a lambda body, but I'd guess that in the error cases it deduces match: Match[bytes] rather than match: Match[str]?

[DMRobertson]

Bisection shows that this was introduced by 7fbf4de ("Sync typeshed"). I'm not sure if this indicates s typeshed bug or just something that a typeshed change happened to expose?

[DMRobertson]

Ahh, digging into the typeshed change (python/typeshed#7679) this looks like this is probably related to (duplicate of?) #12665

[AlexWaygood]

Ahh, digging into the typeshed change (python/typeshed#7679) this looks like this is probably related to (duplicate of?) #12665

Or #12773 (cc. @JukkaL, @JelleZijlstra)

[hauntsaninja]

Yup, duplicate of #12666 / 12665

[hauntsaninja]

The issue is probably something like we're not correctly silencing errors in the earlier passes for the multipass lambda logic