Lifting this restriction will be pretty inconvenient in static type checker implementations — at least in pyright, and I presume in mypy as well. The current implementation simply walks the parse tree to identify top-level statements that correspond to field definitions. If we allow conditional statements, a different approach will be needed — one that involves much more costly code flow analysis to detect unreachable code in conditional branches.

Are we sure the need is sufficient to justify a spec change, all of the downstream work associated with changing this in multiple type checkers, and the performance loss that will result?

The alternative is to leave the current restriction in place and require the version checks to be outside the class body.

if sys.version_info >= (3, 11):
    class A(TypedDict):
        a: int
else:
    class A(TypedDict):
        a: str

The current implementation simply walks to the parse tree to identify top-level statements that correspond to field definitions.

Based on the playground experiments in the question, pyright seems to understand the directives already (at least, in the case with sys.version_info in TypedDict class bodies); it just provides an additional warning on top. With a Python version of 3.13:

import typing as t
import sys

class A(t.TypedDict):
    if sys.version_info >= (3, 11): a: int  # pyright: TypedDict classes can contain only type annotations
    if sys.version_info >= (3, 14): b: str  # pyright: TypedDict classes can contain only type annotations

a: A = {"a": 1,  # OK
        "b": 1}  # pyright: Type "dict[str, int]" is not assignable to declared type "A" \
                 #          "b" is an undefined item in type "A"

It seems like pyright is already walking through conditional statement bodies (not just top-level statements) and correctly determining whether the fields should exist based on the version information, rather than completely ignoring them like mypy is. Do you expect a heavier performance hit to pyright if this idiom were to be used more commonly?

From users' perspective - it is not clear, why if sys.version_info is allowed in regular classes, but not in TypedDicts bodies.

Looking more closely at the implementation in pyright, fields that are under conditional statements that can be evaluated by the binder (which includes version and platform checks documented in the typing spec) would avoid the need for more costly code flow analysis.

I don't know the implementations of other type checkers well enough to speak to whether it would require a significant change or result in a performance hit.

If we do lift this restriction, the spec will need to be more specific about which statements and forms are allowed and which are not.