Expand unit test coverage to 94%, fixing four bugs found along the way

[asjacobs92]

Summary

Coverage was 72% with four modules at 0%. This takes it to 94% — 222 tests, ~70s — and gives every module some.

Module	Before	After
utils/persist.py	0%	92%
utils/rootpath.py	0%	98%
utils/log.py	0%	100%
utils/const.py	0%	100%
report/plot.py	56%	96%
utils/plot.py	84%	95%
report/trust.py	78%	88%
utils/dataset.py	67%	98%
main.py	90%	97%
Total	72%	94%

Bugs the new tests found

Writing tests for the untouched paths turned up four real defects — two of them upgrade fallout that survived 1.2.0 because nothing exercised the code.

np.uint8(-1) raises OverflowError on NumPy 2. It wrapped silently to 255 on NumPy 1, so cic_ids_2017_label_converter() blew up on any unrecognised label instead of returning its sentinel. 255 is now explicit, preserving the old result.

DataFrame.groupby(axis=1) was removed in pandas 2.0, so plot_distribution(aggregate=True) raised TypeError. Rewritten to group bit columns by prefix explicitly, preserving column order so the bits keep their significance. I verified the arithmetic separately rather than just asserting it runs — flags_0/flags_1 fold to [1, 2, 3, 0].

class_names defaults to [] at seven indexing sites guarded only by is not None — which passes for an empty list. Calling any of those entry points without class_names raised IndexError, as did a class_names shorter than the tree's class count. Added a _class_label() helper doing the bounds check trust.py already did.

labels[i] was indexed unguarded in plot_stacked_bars, while every other site in that file bounds-checks.

Behaviour documented, not changed

Three surprises I left alone but pinned with tests, since changing them is a judgement call:

• rootpath.detect() returns None at the first empty directory it meets — an empty subdirectory aborts the search before reaching the marker above it.
• plot_distribution(aggregate=True) requires feature_names; it converts X with .values, so DataFrame column names never reach the prefix regex.
• skip_retrain=True reuses the blackbox as-is and therefore requires an already-fitted model.

Flakiness

Trustee.fit() samples via np.random.choice and calls train_test_split without a random_state — both draw on NumPy's global RNG. My first fidelity assertion passed, then failed on a later run. An autouse fixture now seeds per test; I ran the full suite three times over to confirm it's stable.

Worth flagging separately: the library itself is nondeterministic by default. Trustee(...).fit() gives different explanations across runs unless the caller seeds NumPy globally. That may be intended for the stability analysis, but there's no random_state parameter to pin it. Happy to add one if you want it.

Also

• Every plotting test asserts no matplotlib figure is left open, guarding the leak fixed in 381eb17.
• CI gains --cov-fail-under=90 so coverage can't quietly regress.
• One narrow warning allowlist: joblib 1.5.3 (already latest) sets ndarray.shape directly, which NumPy 2.5 deprecated. It fires on every load_model() and there's no version to upgrade to. Scoped to the exact message so DeprecationWarning stays an error otherwise — worth watching, since it becomes a hard break when NumPy removes it.

[github-actions]

Coverage report

✅ 93.51% meets the 90% floor.

███████████████████░ 1,483 / 1,586 statements · 103 uncovered

File	Coverage	Missed	Uncovered lines
trustee/report/trust.py	88%	71	740, 745, 756, 770–771, 780, 830, +42 more
trustee/utils/persist.py	92%	2	27–28
trustee/utils/plot.py	95%	10	38, 129, 178, 268, 274, 353, +3 more
trustee/report/plot.py	96%	11	232, 522, 570–571, 608–612, 618–619
trustee/main.py	97%	5	217, 269, 509, 536, 553
trustee/utils/rootpath.py	98%	1	104
trustee/utils/dataset.py	98%	2	86, 116
trustee/utils/tree.py	99%	1	28

5 file(s) at 100%

trustee/__init__.py, trustee/_version.py, trustee/enums/feature_type.py, trustee/utils/const.py, trustee/utils/log.py

_{Measured on Python 3.12, ubuntu-latest.}