arch: add chiasm_mode=bias to avoid global-branch starvation

arch-v2.3: confidence-gated output fusion (stabilize ood_long)

arch-v2.2: add QUERY type + soft optic-chiasm split (route query/mid/val to right)

fix: restore lb_loss aux + robust logging; update GradScaler API

arch-v2: asymmetry (local vs global), competitive callosum, optic-chiasm bias; default stem=0

moe: add router warmup (round-robin expert) to reduce seed variance

moe: add load-balance loss + training hook (reduce seed variance)

head: make models query-aware (last-token readout) + trainer update

poc: add bind_query task + english render + side-by-side trainer

probe: add minimal dense + structured models and VRAM probe script