Function fit 

pub fn fit(
    samples: &[TreeSample],
    max_depth: usize,
    min_samples_leaf: usize,
    max_nodes: usize,
    min_rel_var_reduction: f64,
) -> SerializedTree

Grow a CART regression tree over samples.

The growth is breadth-first so the serialized node order is the level-order layout the store requires (parents before children, index 0 = root). Each node is created as a placeholder, queued, and filled when processed. A node that clears the growth limits becomes an internal node (two new children) and everything else becomes a leaf predicting the weighted mean of its labels.

Growth stops at max_depth edges, when either child would fall below min_samples_leaf, when the node budget max_nodes would be exceeded (every node in the tree, internal or leaf, counts), or when no split removes min_rel_var_reduction of the node’s label variance.

Every training sample is weighted by its recency in the window (sample_weights: 2^(-age/(n/2))), so the fit concentrates on the recent regime while the full window still provides the data quantity and the gates bound every publish. The weighting enters the fit as a weighted SSE. The leaf means and the variance-reduction ranking are weighted, and the min_samples_leaf cap still counts samples, not weight.

All arithmetic is f64 on the weighted label sums. The labels are emitted by the BPF side clamped to MLFQ_TREE_LABEL_MAX_NS (see src/bpf/intf.h), so the exact-integer range the SSE math sums is bounded: a label value of at most 192 ms squares to ~3.7e16, far below the f64 rounding error. The weights live in [0.25, 1], so the weighted sums stay within the same magnitude as the uniform fit and the split ranking is exact except for near-tied candidates containing extreme labels, where consecutive u64 values collapse in the f64 conversion and the tie-break is approximate; the daemon never sees such labels because of the emission clamp.

An empty samples slice or max_nodes == 0 yields an empty tree, which serialize_validate() rejects; the daemon treats an empty tree as untrained.