[FIX] Scatter Plot: Update axis labels when discrete values change

[raashish1601]

Issue

Fixes #7308

Description of changes

When new data has the same values in X but a discrete variable's values are renamed (e.g. the steps in the issue: edit the category names in the file and press Reload), Scatter Plot keeps the old category labels on the axis while the legend shows the new ones.

OWDataProjectionWidget.set_data only marks the domain as changed when domain.checksum() differs, and Scatter Plot updates its axes only in that case. The checksum is the domain's hash, and a variable's hash is built from its name, type and compute value, not its values. So renaming the values leaves _domain_invalidated false, the data is not invalidated either (the arrays are equal), and update_axes is never called.

The domain is now also treated as changed when the values of its discrete variables differ.

Tests: test_discrete_axis_values_changed sends data with a discrete variable valued ("1", "2"), then the same table with values ("a", "b"), and checks the x-axis ticks. It fails without the change (the ticks stay 1, 2) and passes with it. test_owscatterplot.py, test_owlinearprojection.py, test_owradviz.py, test_owfreeviz.py and test_owmds.py pass locally (Windows, against the Orange 3.40 wheel with this patch applied, apart from one test in the master copy of test_owscatterplot.py that needs a newer graph than 3.40 has).

Includes

• Code changes
• Tests
• Documentation

[CLAassistant]

All committers have signed the CLA.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 88.97%. Comparing base (eea7a75) to head (cd44b04).

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #7336      +/-   ##
==========================================
- Coverage   88.98%   88.97%   -0.01%     
==========================================
  Files         337      337              
  Lines       74598    74601       +3     
==========================================
+ Hits        66378    66379       +1     
- Misses       8220     8222       +2     

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[raashish1601]

The "Scientific Python nightly wheels" failure is unrelated to this change: the nightly pyqtgraph turned LegendItem.layout into a method, so hundreds of widget tests error out. Master fails the same way (https://github.com/biolab/orange3/actions/runs/37914329734).

[janezd]

@raashish1601, thanks. The failed test is indeed caused by an (irresponsible) change in pyqtgraph and would be patched in #7329.

The diagnosis and the fix is essentially correct, but I think that it would be better to put the change deeper.

The genesis of the problem is such.

• Scatter plot uses Domain.checksum to check whether the domain has changed.
• Domain.checksum returns hash(domain).
• Hash of a domain is composed of hashes of variables.
• Hash of a variable takes into account its name, type and compute value -- but not a list of values of discrete variable.

We can fix the code in three places:

• DiscreteVariable.__hash__ should include values. Yet, this hash is used all around the code, so the change might break something. (Or, possibly, fix something.) This would be a proper fix.
• Domain.checksum could call Domain's hash and add values. This is strange, but safe: it can't break anything (except possibly in add-ons) because this method is used called exclusively from the scatter plot.
• Current PR, which changes the scatter plot by checking discrete values. This is safest and not strange, but partial -- essentially a very late patch.

I would lean towards DiscreteVariable.__hash__ and hope that tests will discover any code that requires the current (invalid) implementation.

@markotoplak, @VesnaT, what do you think?

[raashish1601]

Thanks for the explanation. I'm happy to move the fix into DiscreteVariable.__hash__ (with a test there) if that's the direction you pick, and drop the scatter plot change. I'll wait for the decision.