MyPy treats initialization of a variable as an assignment for variance checks for complex dictionary.

[markshannon]

The problem is that when initializing a variable from a dictionary expression it treats the type as invariant, but in the case of initialialization, it is covariant.

Example:

CODES = {
    "a": "b",
    "c": "d",
    "e": "f",
}

# MyPy rejects this

PATTERNS: dict[str, tuple[str | None, str | None]] = {
    "1" + k: ((v + "1" if v else None), "x")
    for (k, v) in CODES.items()
} | {
    "2" + k: ((v + "2" if v else None), "x")
    for (k, v) in CODES.items()
}

# But is OK with this:

X: str | None = "x"

PATTERNS2: dict[str, tuple[str | None, str | None]] = {
    "1" + k: ((v + "1" if v else None), X)
    for (k, v) in CODES.items()
} | {
    "1" + k: ((v + "1" if v else None), X)
    for (k, v) in CODES.items()
}

Both patterns should be OK.

[A5rocks]

Probably we're forgetting to pass through context to a in a | b. This might be intentional because those rules are mostly based on heuristics for better errors...

[hauntsaninja]

Related: python/typing#2104

[A5rocks]

Actually yeah we only know the return type of __or__ (or whatever the dunder is). I guess we could synthesize a protocol and then infer against it?