fix(trees): make DecisionTree::fit deterministic

[mysma-9403]

DecisionTree::fit is not reproducible: repeated calls on identical data return differently shaped trees.

On the iris dataset, 60 fits in a single process produce 4 distinct trees, differing in leaf count (9, 10, 11, 12) and therefore in their predictions:

L[Some(0)@1,Some(1)@3,Some(2)@3,Some(2)@4,...]   (11 leaves)
L[Some(0)@1,Some(2)@2,Some(1)@3,Some(2)@4,...]   ( 9 leaves)
L[Some(0)@1,Some(2)@2,Some(2)@4,Some(1)@4,...]   (10 leaves)
L[Some(0)@1,Some(2)@3,Some(2)@4,Some(1)@4,...]   (12 leaves)

Cause

Class frequencies were carried in a HashMap<L, f32>, and HashMap iteration order is randomised per instance. Both consumers of those frequencies depend on the order:

1. find_modal_class folded with if best_freq > freq { acc } else { Some((idx, freq)) }, so on a tie it kept the last maximum it happened to visit — the opposite of the "first class found with that frequency is returned" its own doc comment promises.

2. gini_impurity and entropy sum the frequencies in f32, and floating-point addition is not associative. The same frequencies summed in a different order give slightly different scores.

The second one is what changes the tree shape: those small score differences flip the score < best_score comparison in TreeNode::fit at near-ties, selecting a different split and changing the whole subtree below it.

Fix

Frequencies become Vec<f32> indexed by position in a sorted class list (ClassIndex) built once per fit, so iteration order is pinned. find_modal_class now also uses >= so that ties resolve to the first class, matching its documentation.

The tree this produces is one of the trees the previous code already produced — I verified this by collecting the fingerprint set of 60 fits on master and checking that the now-deterministic output is a member of it. This changes which of several equally-valid trees you get, consistently, rather than introducing a new one.

Side effect: it is also faster

The innermost split-search loop ran, for every (sample, feature) pair:

*right_class_freq.get_mut(sample_class).unwrap() -= sample_weight;
*left_class_freq.entry(sample_class.clone()).or_insert(0.0) += sample_weight;

Two hash lookups plus a label clone — for String labels, a heap allocation. Indexing a Vec removes all three.

n	d	classes	before	after	speedup
2 000	20	3	707.2 ms	271.5 ms	2.6×
8 000	20	3	2570.1 ms	860.9 ms	3.0×
8 000	20	10	2198.7 ms	861.1 ms	2.6×
16 000	50	5	14652.5 ms	4494.8 ms	3.3×

Benchmark context

1. Plugged in (AC power, battery 100 %, charged)
2. Power saving mode: off (lowpowermode 0)
3. Machine otherwise idle
4. Not thermally throttled
5. MacBookPro15,3 — Intel Core i9-9980HK @ 2.40 GHz, 8 cores / 16 threads, 32 GB RAM, macOS 15.7.7, rustc 1.93.1
6. Best of 5 runs of DecisionTree::params().fit(&ds), default hyperparameters, measured with a throwaway harness (not part of this PR); A/B taken on this branch with and without the change so the binary shape is identical

Notes

• ClassIndex::frequencies replaces DatasetBase::label_frequencies_with_mask inside the tree. The only behavioural difference is that classes absent from a node are present with frequency 0.0 instead of being omitted, which is inert: a zero frequency contributes nothing to either impurity measure, and can never be the modal class of a non-empty node.
• Added fit_is_deterministic as a regression test, plus modal_class_ties_pick_the_first_class to pin the documented tie-break.
• The three existing unit tests for gini_impurity/entropy/find_modal_class were updated for the new slice signatures; their expected values are unchanged.

Checks

• cargo fmt --all -- --check — clean
• cargo clippy --workspace --all-targets -- -D warnings — clean
• cargo test --release --workspace — 546 tests, no failures

🤖 Generated with Claude Code

https://claude.ai/code/session_01L4uUK6s1CWFwx6u7Rx28go

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 78.04%. Comparing base (7fe5c86) to head (4a52706).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #452      +/-   ##
==========================================
+ Coverage   77.53%   78.04%   +0.51%     
==========================================
  Files         106      104       -2     
  Lines        7585     7543      -42     
==========================================
+ Hits         5881     5887       +6     
+ Misses       1704     1656      -48     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.