Speed up HalfSpaceTrees with an iterative walk

[MaxHalford]

Summary

Replaces the recursive generic tree.base.Branch.walk traversal in anomaly.HalfSpaceTrees.learn_one/score_one with HST-specific iterative loops that:

• Inline the split-and-descend logic (no per-level generator allocation, no HSTBranch.next call).
• Cache size_limit = 0.1 * window_size and self.height as locals (was a property recomputed ~1.1M times in profile).
• Pivot node masses through a precomputed flat node list built once at tree construction, instead of iter_dfs() every window_size observations.

The public API is unchanged — no constructor changes, no behaviour changes. The docstring scores remain byte-identical and all 50 check_estimator cases for MinMaxScaler | HalfSpaceTrees pass.

Benchmarks

macOS arm64, Python 3.13, default HalfSpaceTrees(n_trees=10, height=8, window_size=250), best of 3.

Pure HST on a synthetic 10-feature uniform stream (no scaler):

Workload	Before	After	Speedup
score+learn	27,902 obs/s	85,231 obs/s	3.05x
learn_one	64,232 obs/s	169,098 obs/s	2.63x
score_one	49,598 obs/s	188,172 obs/s	3.79x

MinMaxScaler | HalfSpaceTrees on datasets.CreditCard() (10k samples):

Workload	Before	After	Speedup
score+learn	20,493 obs/s	40,513 obs/s	1.98x
learn_one	42,943 obs/s	74,241 obs/s	1.73x
score_one	38,590 obs/s	89,623 obs/s	2.32x

In the post-fix profile, HST itself (_walk_learn + _walk_score) is no longer the dominant cost — MinMaxScaler.transform_one/learn_one is. Further wins would need a Rust port (cf. the recent ADWIN/Mondrian/VectorDict efforts) or speeding up MinMaxScaler.

Test plan

• uv run pytest --doctest-modules river/anomaly/hst.py — passes, docstring outputs byte-identical
• uv run pytest river/anomaly/ -x — all 16 tests pass
• uv run pytest river/test_estimators.py -k "HalfSpaceTrees" — all 50 estimator checks pass
• pre-commit (ruff, mypy) — passes

🤖 Generated with Claude Code