Flaky test: test_dump_modules[description_no_llm] — CategoricalDistribution does not support dynamic value space

[voorhs]

Symptom

tests/pipeline/test_optimization.py::test_dump_modules[description_no_llm] fails intermittently with:

ValueError: CategoricalDistribution does not support dynamic value space.

The test sometimes passes, sometimes fails, on identical commits — i.e. it is a real flake, not a regression.

Where it surfaces

• dev branch: alternating green/red across consecutive CI runs (e.g. run 27102499946 green, prior run 27101359057 red, run before that 27098583531 green).
• mypy-on-tests Phase B: surfaced on two parallel subagent PRs (test(types): annotate tests/modules/decision (51→0 mypy errors) #304, test(types): annotate tests/modules/{embedding,test_dumper,test_regex} (45→0 mypy errors) #307) whose diffs do not touch tests/pipeline/ at all, confirming it's not driven by the diff.

Example failing job logs:

• https://github.com/deeppavlov/AutoIntent/actions/runs/27127958767/job/80061423953 (PR test(types): annotate tests/modules/decision (51→0 mypy errors) #304)
• https://github.com/deeppavlov/AutoIntent/actions/runs/27128038542/job/80061871290 (PR test(types): annotate tests/modules/{embedding,test_dumper,test_regex} (45→0 mypy errors) #307)

Likely root cause

The failure is preceded by warnings emitted from src/autointent/nodes/_node_optimizer.py lines 386, 389, 392, 394:

UserWarning: Choices for a categorical distribution should be a tuple of
None, bool, int, float and str for persistent storage but contains
{'model_name': 'sergeyzh/rubert-tiny-turbo', 'device': 'cpu'} which is of type dict.

Optuna requires CategoricalDistribution choices to be hashable scalars for persistent storage. The search-space currently passes dict objects (embedder configs) as choices, which Optuna stores via repr() — when the persisted study is loaded back, the "value space" reconstructed from string reprs of dicts no longer matches by identity, and the description_no_llm variant's particular ordering of choices appears to trip the equality check non-deterministically (insertion order in Python dicts is stable but dict repr ordering interacts with set/list de-dup downstream in Optuna).

The flake-ness comes from this trip being sensitive to ordering decisions inside Optuna's storage layer, which can shift between runs.

Suggested fix direction

Convert dict-typed categorical choices into a hashable, scalar representation before handing them to Optuna — e.g. a stable string key ("st:sergeyzh/rubert-tiny-turbo:cpu") or an index into a side-table of full configs. Resolve the chosen key back to the dict at trial time.

This would:

1. Eliminate the four UserWarnings pointing at _node_optimizer.py:386-394.
2. Make the persisted study round-trippable cleanly.
3. Remove the flake.

Workaround

Rerun the failing job. The test is unrelated to the bulk of diffs flagged in CI for it.

Discovered during

The strict-mypy-on-tests workstream (Phase B fan-out). Not blocking that workstream — the failing test is outside every subagent's diff.