[Copilot]

A test for variational_gamma was skipped pending a decision on whether it should raise NotImplementedError when a sample has undated descendant nodes. The EP machinery turns out to handle this case already, so no error is needed.

• Test change only (tests/test_functions.py)
  • Removed the skipped test_variational_gamma_sample_as_parent_fails.
  • Added test_variational_gamma_sample_as_parent. It adds a mutation to single_tree_ts_n3_sample_as_parent(), because the fixture has none and variational_gamma requires mutations. It then checks that sample ages are unchanged after dating.
• No library change
  • A guard that raised on samples-as-parents broke TestInternalConstraints::test_with_unary_internal_samples and test_with_nonunary_internal_samples. Both already date internal samples through date().
  • A guard on non-time-0 samples broke the same two tests, so non-contemporary samples are also supported. I reverted both guards.

ts = single_tree_ts_n3_sample_as_parent()  # sample 3 is the parent of samples 0 and 1
dts = tsdate.variational_gamma(ts, mutation_rate=1, rescaling_intervals=1)
# nodes_time: [0, 0, 0, 1, 1.00000001]; sample 3 stays at 1

Review notes:

• The new test passes rescaling_intervals=1. With the default, this three-sample fixture fails with AssertionError: Use fewer rescaling intervals. That is a limit of piecewise rescaling on a tiny tree, not a sample-as-parent problem, and I didn't change it.
• The coverage is a single three-sample tree with one mutation. It checks that sample ages are preserved, not the quality of the inferred ages for the non-sample nodes.

• Fixes sample_as_parent in variational_gamma #394

[hyanwong]

This LGTM. The functionality was added in #433 (there is a lot of discussion and some more test cases there)

[hyanwong]

This just reactivates a test that should now pass, and is uncontroversial