fix: isinstance narrowing regression with dynamic tuple argument

[Bahtya]

Problem

When isinstance is called with a dynamically-computed tuple of types (e.g. isinstance(exc, tuple(expected_excs))), mypy 1.20 incorrectly narrows the expression to object instead of keeping its already-narrowed type. This is a regression from 1.19.

def f(exc: BaseException | None, expected: set[type[BaseException]]) -> None:
    if exc is not None:
        if isinstance(exc, tuple(expected)):
            reveal_type(exc)  # 1.19: BaseException ✅ | 1.20: object ❌

Root Cause

The second argument to isinstance has its type stored as the expanded _ClassInfo recursive type alias (type | types.UnionType | tuple[_ClassInfo, ...]) rather than the precise tuple[type[BaseException], ...].

In get_type_range_of_type, the UnionType handler decomposes this union into individual members. The bare type member produces TypeRange(object, is_upper_bound=True), which after union simplification becomes the sole result. This causes conditional_types to narrow to object — a broader type than the already-narrowed BaseException.

In mypy 1.19, the raw UnionType was not handled by the type-range logic (it fell through to return None), which correctly preserved the existing narrowed type.

Solution

In get_type_range_of_type UnionType handler:

1. Propagate None if any sub-item returns None (uncertainty)
2. Filter out UninhabitedType entries before simplifying
3. If the simplified result is just object, return None — this indicates type precision was lost (e.g. from the _ClassInfo widening), and we should fall back to keeping the current type

Testing

• Verified fix against the original issue reproduction
• All 248 isinstance tests pass
• All 209 narrowing tests pass
• All 843 optional/tuple tests pass
• No regressions observed

Fixes #21181

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

psycopg (https://github.com/psycopg/psycopg)
- psycopg/psycopg/types/numeric.py:407: error: No overload variant of "int" matches argument type "object"  [call-overload]
- psycopg/psycopg/types/numeric.py:407: note: Possible overload variants:
- psycopg/psycopg/types/numeric.py:407: note:     def int(str | Buffer | SupportsInt | SupportsIndex = ..., /) -> int
- psycopg/psycopg/types/numeric.py:407: note:     def int(str | bytes | bytearray, /, base: SupportsIndex) -> int

hydra-zen (https://github.com/mit-ll-responsible-ai/hydra-zen)
+ src/hydra_zen/_launch.py:92: error: Redundant cast to "T"  [redundant-cast]

[Bahtya]

Hi mypy maintainers! Friendly ping on this PR.

To summarize for ease of review:

• All CI checks pass (including mypy_primer, mypyc, self-check across Python 3.10–3.14)
• mypy_primer result: The fix correctly eliminates a false call-overload error in psycopg, which was caused by the regression this PR addresses. The one new note in hydra-zen (redundant-cast) is a pre-existing unrelated issue.
• The fix is surgical: it only touches the UnionType handling in get_type_range_of_type, with no risk of over-narrowing elsewhere.
• Fixes regression [1.20 regression] when same object is narrowed multiple times #21181 (isinstance narrowing became overly broad with dynamic tuple args in mypy 1.20).

Happy to make any adjustments if needed. Thanks for your time!