Fix flaky Hypothesis stateful test for unique name generation

[max-sixty]

Claude tried to fix the flakiness issue in CI. I probed its logic a bit, it looks reasonable but I don't fully understand. Our summary below...

---

Summary

• Fix flaky DatasetStateMachine test that was causing CI failures with FlakyStrategyDefinition: Inconsistent data generation!
• Move unique name tracking from st.shared() strategy to state machine instance state

Problem

The nightly Slow Hypothesis CI has been failing:
https://github.com/pydata/xarray/actions/runs/19603404994

The test properties/test_index_manipulation.py::DatasetTest::runTest fails intermittently with:
hypothesis.errors.FlakyStrategyDefinition: Inconsistent data generation!
Data generation behaved differently between different runs.
Is your data generation depending on external state?

Root Cause

The original code used a pattern from [Stack Overflow](https://stackoverflow.com/questions/73737073/c
reate-hypothesis-strategy-that-returns-unique-values):

@st.composite
def unique(draw, strategy):
    seen = draw(st.shared(st.builds(set), key="key-for-unique-elems"))
    return draw(
        strategy.filter(lambda x: x not in seen).map(lambda x: seen.add(x) or x)
    )

While this pattern works for regular @given tests, it does not work for RuleBasedStateMachine
because:

• The shared set's contents depend on which rules ran before
• During shrinking, Hypothesis replays with different rule orderings
• This causes different names to be drawn, breaking determinism

Per HypothesisWorks/hypothesis#2338:
"I think you'll need to set these up as instance attributes, rather than class attributes."

Solution

Move uniqueness tracking to state machine instance state:

• Added self.used_names: set[str] (reset per test case via init)
• Added _draw_unique_name() that filters and tracks used names
• Updated all rules to use data=st.data() and call helper methods

This follows the Hypothesis-recommended pattern for stateful tests where cross-rule state belongs on
the instance.

[dcherian]

Thank you! I looked at it a couple of months ago and couldn't figure out what was going on