mo.state setters should not cause an update to the cells with getters

[umnikos]

Currently in a Marimo notebook when a cell calls a mo.state setter, all of the cells with getters (other than the one who called the setter) get re-ran. This is problematic for two reasons:

1. This dependency is not reflected in the DAG; As far as the DAG is concerned both the getter and the setter cells just depend on the cell that defined the mo.state, there is no dependency from the setters to the getters
2. This behavior is inconsistent; When that very same notebook is embedded through app.embed(), setting a mo.state no longer causes an update to all of the cells with getters.

Because of these problems mo.state has very limited use. You cannot rely on the setter to update the getters and you cannot rely on the setter to not update the getters. It cannot be used to get rid of dependency cycles in the DAG because if the setter updates the getters then the cycle still exists, it's just that instead of Marimo catching the cycle and erroring you get an infinitely looping notebook. You also cannot use mo.state to control python code from widgets because the mechanism breaks when the notebook gets embedded.

Due to the existence of traitlets for the latter use, I propose we make mo.state handle the former use. This change will not only make mo.state more useful and consistent in behavior, but also make its semantics properly reflected by the DAG.

[Light2Dark]

This behavior is inconsistent; When that very same notebook is embedded through app.embed(), setting a mo.state no longer causes an update to all of the cells with getters.

If this is a bug, then we should fix it. We do discourage using mo.state, but it has its uses.

[umnikos]

If this is a bug, then we should fix it.

Currently it's a bug, yes. I propose that instead of the embedded behavior being changed the normal behavior should be changed.

[Light2Dark]

Thanks. Could you produce a minimum code example, and we'll look into it? Can convert this to an issue.

[umnikos]

...I just realized a version of mo.state that doesn't update the getters already exists and is called a python dictionary, so no changes are necessary to mo.state's design.

I'll make a bug report for the embedding bug with a minimal code example when I get the chance (next week probably) since it seems that more than just app.embed is required to trigger the bug.