Implement the iforest method of IsoKernel

[xhan97]

The bug

IsoKernel documents iforest alongside anne and inne, and dispatches to it — but IK_IForest was a docstring with def __init__(self): pass underneath. Following the documentation crashed:

>>> IsoKernel(method="iforest").fit(X)
TypeError: IK_IForest.__init__() takes 1 positional argument but 4 were given

The implementation

Each estimator draws max_samples points and grows an isolation tree over them — one feature at a time, random threshold, to the height a standard isolation forest stops at (ceil(log2(ψ))). A point's feature is the leaf it lands in, so the cells are axis-parallel boxes, where anne has Voronoi cells and inne has hyperspheres.

Built from ExtraTreeRegressor directly rather than IsolationForest, which would also fit an anomaly scorer (offset_) over all of X that this has no use for. The structure mirrors _ik_anne.py: per-estimator seeds, per-estimator subsample.

One thing worth knowing

A tree numbers its leaves by their position in its node array, so the ids run past the number of leaves and can't be used as feature columns:

tree 0: node_count=15  distinct leaves=8   max leaf id=14
tree 2: node_count=19  distinct leaves=10  max leaf id=18

They're mapped onto a contiguous range. That range always fits the block width the other two methods use — a tree grown on ψ points has at most ψ leaves, because every leaf holds at least one of them — so transform keeps the existing (n_samples, n_estimators × max_samples) shape contract that test_IsoKernel_transform already asserts. Verified empirically across ψ from 4 to 256; leaf counts topped out at ψ every time.

Tests

method in the test module was ["inne", "anne"] — iforest was excluded because it didn't work. It's now included, so the three existing tests cover it, plus three new cases asserting what the kernel is supposed to guarantee:

Test	Checks
..._is_one_hot_per_estimator	Exactly one cell per estimator. inne is excluded from the lower bound — a point outside every hypersphere legitimately falls in no cell
..._is_reproducible	Same random_state → identical embedding; different → different
..._similarity_is_higher_in_sparse_regions	The property the kernel exists for: two neighbours alone in a sparse region score as more similar than two neighbours inside a crowd

On that last fixture:

method	sparse pair	dense pair
anne	0.977	0.110
inne	0.153	0.003
iforest	0.970	0.227

Also added a case for the unknown-method ValueError, which had no coverage.

Scope

IDKD, IDKC, IKAHC and ICID each whitelist only the methods their papers specify — IKAHC rejects anything outside ["inne", "anne"] explicitly. Widening those changes published algorithm behaviour and is a separate decision, so they're untouched.

Verification

ikpykit/kernel/tests/ — 25 passed locally. The full suite is left to CI on this PR; local runs are slow here. All pre-commit hooks pass.

🤖 Generated with Claude Code