Skip range_extension_c for pure nodes in Mondrian classifier

[MaxHalford]

Summary

• When split_pure=False (default), moves the purity check before the range_extension_c call in the classifier's downward pass. Pure nodes never split, so the range extension computation can be skipped entirely.
• Credit to @shi-zq for the observation.

Benchmark results

range_extension_c cost scales linearly with feature count (~1μs at 20 features, ~10μs at 200).

Pure nodes are ~50% of all nodes (exclusively leaves — branches are always impure). The optimization saves one range_extension_c call per learn_one at the terminal pure leaf.

Config	Baseline	Optimized	Delta
20 features, 3 classes	0.1781s	0.1774s	-0.4%
50 features, 3 classes	0.4067s	0.3876s	-4.7%
100 features, 3 classes	0.6655s	0.6425s	-3.5%
200 features, 3 classes	1.0568s	1.0203s	-3.5%
50 features, 2 classes	0.4338s	0.4128s	-4.8%
50 features, 5 classes	0.4071s	0.3891s	-4.4%
50 features, 10 classes	0.3868s	0.3692s	-4.6%

(5000 samples, best of 5 runs, split_pure=False, seed=42)

Test plan

• pytest river/tree/mondrian/ passes (including doctests)
• CI green

🤖 Generated with Claude Code