Exact leaf sums and groups of any width

[ukaratay]

predict() now returns the correctly rounded sum of the leaf values and accepts groups of any number of rows. The first four commits add the research kernels; the rest measure them, fold what paid into predict(), and remove the kernels again.

The native XGBoost baseline never predicted. Its config lacked missing and the iteration fields, so XGBoosterPredictFromDense returned -1 on every call ("Argument missing is required", XGBoost 3.2.0, checked 2026-10-03), and xgboost_native timed the error path: a flat ~180 µs at any model size. The released CSVs are flat the same way (289–295 µs Intel, 208–240 µs Arm, across all grid cells), so the XGBoost-native comparisons in paper/experiments/scripts/paper_numbers.py (lines 103, 116, 168–170) need a rerun. The harness now asserts the return code.

Leaves are summed exactly. Every leaf value of a model is an integer multiple of 2^-e for one scale e; a leaf adds v·2^e as an i128 to a difference array once per run of rows in its mask, and each row's prefix sum is rounded to f64 once. A leaf costs O(runs) instead of O(rows), and a prediction no longer depends on tree order, node layout, ablation mode or group width. A model with a nonfinite leaf, or with leaf exponents too far apart for 126 bits, adds in f64 in tree order instead; Forest::exact_sums() reports which.

Masks are u16, u32, u64, or Bits<W> of 64-bit words up to 1,024 rows. max_group_width can be any positive number, and wider groups run in 1,024-row pieces; a row predicted in a group of any width equals the row predicted alone, bit for bit. predict() also compiles out the ablation checks, which only predict_with_stats() reads.

On an M3 Pro (2026-10-03, 500 trees of depth 8), the new predict() is 1.11–1.13× faster than the previous f64 accumulation at 16 rows, 1.6–1.7× at 104–128 rows, and unchanged on Expedia sessions. Groups of 256–1,008 rows run 17–31% faster in one pass than in 128-row chunks. The run-list kernel was faster on monotone panels (1.2× at 256 rows, 2× at 1,008) but needs shape checks and a fallback path, so it is not included. Child-kind hints, 0–7% on a pinned Xeon, wait for the full suite run, which these numbers do not replace.

Breaking, so this needs a major version before publishing:

• predict::MAX_GROUP_WIDTH is removed.
• RowMask has new required methods and no u128 impl.
• LightGBM predictions move in the last bits: up to 2.2e-15 on the four survival cells measured.

Tests:

• A fixture test checks groups of 129–2,500 rows against single rows bit for bit; another checks that parse-time layouts cannot change a prediction.
• The artifact suite's LightGBM tolerance against GTIL goes from 1e-14 to 1e-13. On the 1,008-step FLCHAIN panel, where margins reach ±500, GTIL's own f64 rounding is up to 1.06e-14 from the correctly rounded sum (math.fsum over the per-tree outputs, all 1,314,432 rows).
• The ablation test used to call predict(), which ignores ablation. It now runs the ablations, skipping the monotone-scan modes on the tiled chunked-G cells, which break the monotonicity contract by construction.
• test_binary_matches_json fails on the Expedia LightGBM artifact: Treelite's JSON dump writes "threshold": , for nonfinite thresholds, which the loader rejects as docs/treelite-loading.md documents.