Fix parallel-call eval scoring, Muon update, and assorted correctness issues

[Joe-Khawand]

First time contributing here — spent some time in the training and inference paths and put together patches for a few things that looked like bugs. Happy to split them if that helps review.

Evaluation (training/eval.py)

• Call P/R used set intersection, so identical parallel calls collapsed (emitting one of two identical required calls scored full recall). Now multiset-counted.
• Arg/value metrics always compared against ref_by_name[name][0], mis-scoring parallel_multiple. Now greedy bipartite matching to distinct reference instances.
• Added BFCL-style category buckets (simple / multiple / parallel / parallel_multiple) plus irrelevance, relevance and false-trigger rates. Scoring split into a pure score_tool_calls().

Training

• Muon (optim.py): the raw gradient was orthogonalized before momentum, so the applied update was a sum of orthogonal matrices — not itself semi-orthogonal. Now momentum runs on the raw gradient and the Nesterov-blended update is orthogonalized; added max(1, fan_out/fan_in)^0.5 scaling. This shifts update magnitude, so muon_lr will want re-tuning.
• z-loss (train.py, pretrain.py): averaged over padding while CE is masked, so its weight drifted with the padding fraction. Now masked to real tokens.

Inference (model/run.py, model/architecture.py)

• decode() can project only the current position instead of the full (B, T, d) @ (d, V) buffer every step — numerically identical, training path unchanged. Projection only; a KV cache is the bigger win and left for a follow-up.
• Streaming decoded each token in isolation, dropping SentencePiece spaces and splitting multibyte byte-fallback pieces. Now emits the decode delta.

Data (dataset/)

• Loss mask marked only the first occurrence of each name/value, leaving repeats across parallel calls at base weight. Now marks all.
• Boolean-polarity validator matched substrings ("on" inside "location", "song", …), which disabled the inverted-boolean filter. Now whole-word / phrase.

Validation

No access to the training data, so this is validated without a training run: the pure-Python logic (eval scoring, Muon ordering, polarity, loss mask) has unit tests, and the inference changes are checked against needle.pkl (decode(cur_pos=k) is bit-identical to the full projection; generation unchanged). The training effects of the Muon/z-loss and data changes aren't validated — they'll want an LR sweep / regeneration on your side.

Glad to take feedback and adjust anything — and thanks for open-sourcing this.