fix: fix name collisions between dataclass fields and guiclass container attributes

[tlambert03]

fixes #684

When a container has a widget whose name attribute matches an attribute on the Widget class itself, then the overloaded __getattr__ method becomes a problem. (reminder: that getattr is what makes it possible to retrieve a magicgui function parameter (for example) as an attribute on the magicgui container widget).

Here I add a get_widget(self, name:str) method on Container that is guaranteed to either return a widget instance or None, avoiding conflicts with other attributes no the Widget instance itself

[codecov]

Codecov Report

Attention: Patch coverage is 86.76471% with 9 lines in your changes missing coverage. Please review.

Project coverage is 88.85%. Comparing base (e18ddc6) to head (23b38e1).

Files with missing lines	Patch %	Lines
src/magicgui/widgets/bases/_named_list.py	88.23%	4 Missing ⚠️
src/magicgui/schema/_guiclass.py	83.33%	3 Missing ⚠️
src/magicgui/experimental.py	0.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #688      +/-   ##
==========================================
- Coverage   88.86%   88.85%   -0.02%     
==========================================
  Files          39       40       +1     
  Lines        4761     4809      +48     
==========================================
+ Hits         4231     4273      +42     
- Misses        530      536       +6     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[tlambert03]

@hanjinliu, could you have a look at the magic-class error and see what i've done here that is selectively breaking magic-class? thanks!

[hanjinliu]

Hi @tlambert03 , I found that the older version and this version returns inconsistent result when a Container widget has multiple widgets with the same name.

from magicgui import __version__
from magicgui.widgets import Container, LineEdit
c = Container()
c.append(LineEdit(name="a", value="1"))
c.append(LineEdit(name="a", value="2"))
c["a"].value

This code returns 1 in the older version, but returns 2 for this PR. The magicclass error occurred because there are widgets of different types.
In the order version, __getattr__ (and __getitem__) searches for the widget whose name matches the input str by for-loop from the beginning of _list. Maybe something changed in this PR?

[tlambert03]

ok, that makes sense... it's caused by this change:

https://github.com/tlambert03/magicgui/blob/8deddf6c7e93cb160ca5dbb06ded2e666a2160c5/src/magicgui/widgets/bases/_named_list.py#L26-L36

I'd say that the implicit expectation of getting the first matching widget in the case of a name collision rather than the last matching widget was probably always more bug-like and not any sort of formal protocol here, but if that's all it takes not to break magic-class I can go back to that behavior. Perhaps after this PR, you can also have a look at places where you have behavior (either explicit or implicit) that might change depending on whether someone has added two widgets with the same name attribute. I don't think either of us should depend on name lookup in that case...

In the case of a dataclass (which precipitated this PR), i think it is safe to assume that there will never be two widgets with the same name, but that is probably a unique case

[hanjinliu]

Yes, I definitely have to avoid adding widget of the same name on my side.
I think that adding widgets of the same name is always dangerous and it is a good idea to encourage people to make widgets from dataclass, TypedDict, pydantic BaseModel etc. to prevent collision errors at the language level.