Thinking-aware retrain: 4 epochs on 28.6k corpus (45% thinking traces); val cond_acc 0.622 -> 0.685/0.697/0.715. Thinking-mode acceptance 0.33->0.41 (+23%), thinking throughput 244->272 tok/s. Card, chart, val_metrics, train_command updated.

Embed long-doc spec chart

Add long-doc spec comparison chart (real data)

Correction: long-document summarize measured on real data favors ngram (231 vs 160); head wins short-context workloads

EAGLE-3 draft model for Nanbeige4.2-3B: epoch-3 head, looped-arch plugin fixes, training curves

initial commit