fix: handle PEP 604 unions (X | Y) like Union[X, Y] - #746

Follow-up to #745, which uncovered this.

The bug

get_origin reports a different origin for the two union spellings on python < 3.14:

get_origin(Union[int, None]) is Union   # True
get_origin(int | None)       is Union   # False on 3.11-3.13, True on 3.14

magicgui compared get_origin(...) is Union in four places, so every PEP 604 annotation quietly took the non-union path. Two user-visible consequences:

@magicgui
def f(x: Optional[int] = None): ...   # f.x.annotation -> int          ✅
@magicgui
def f(x: int | None = None): ...      # f.x.annotation -> int | None   ❌

register_type(int | str, return_callback=cb)        # registered nothing for int or str
register_type(Union[int, str], return_callback=cb)  # registers both  ✅

Widget selection was unaffected — _split_annotation_type and _literal_choices use get_args() directly, which works for both spellings. Only the four is Union comparisons were wrong, which is why this stayed hidden. Masked entirely on 3.14, where types.UnionType became typing.Union.

The fix

Adds magicgui._util.is_union, which accepts both spellings, and uses it at all four sites (_type_map, _ui_field ×2, _value_widget).

Re-enabling the pyupgrade rewrite

#745 put UP007/UP045 in ruff's ignore list because of the bug above. With it fixed, UP045 is safe and is now enabled — but scoped deliberately:

• UP045 on for src/ and docs/, off for tests/. Tests exercise both spellings on purpose; blanket-rewriting them would cost coverage (this is what neutered test_no_order in ci(pre-commit.ci): autoupdate #728).
• UP007 stays off. Union is still required as a runtime value for the public type aliases (PathLike, ChoicesType, AppRef, TableData, WidgetRef) and for Union[args] construction. ruff offers no fix for those 11 sites — one of them, WidgetClass, holds string forward refs where | would be a runtime TypeError — so enabling it would only leave permanent lint errors.

The rewrite itself is small: 3 files, all genuine annotation positions. docs/examples/demo_widgets/optional.py now uses str | None, which means test_examples.py exercises the fix end to end.

Also included: a lint fix for main

main is currently failing ruff. #742 added from typing import Callable to the ipynb backend and merged after #745 switched target-version to py311, so its CI ran against the old config and never saw UP035.

Related issues

I searched the tracker; none of these report this bug, and this PR does not close any of them. Listing them because they're adjacent and were candidates:

• Supporting Union #93 — feature request for swappable Union widgets, unrelated
• Optional types should map to the same widgets as non-optional and be bound to the default None #145 / After update to 0.7.0 magicgui stop handling optional return #541 — Optional widget-mapping and return-handling, both typing.Optional
• guiclass throws type error with Optional argument set to None #706 — guiclass with Optional[int] = None. Still fails after this PR (TypeError: float() argument must be...); it's a value-coercion bug, not an origin bug. What this PR does guarantee is that Optional[int] and int | None now fail identically there — previously they diverged.

Testing

Two regression tests in tests/test_types.py, both confirmed failing before the fix and passing after. Full suite green on 3.11 / 3.13 / 3.14 across PyQt6, PyQt5, PySide6 (439 passed).

🤖 Generated with Claude Code