refiner: ReZero gates (a_attn/a_ffn init 0) -> identity at init, stable deep weight-shared iteration (fixes eval-BPB divergence)

GLR-1 refiner: PatchRefiner (weight-shared iterative patch attention) + deep-supervision/PonderNet losses; end-to-end wired, causal, tested

Section 10: sparse executable operator memory (rank-1 pool, top-k router, KG-05 controls)

KG-04 fixes: orthogonal factor_emb (de-redundancy) + optional local attn

mixed-R kernel: exact vectorized patch-expansion chunk scan for R>1/adaptive (2.4x CPU, far more on GPU)

KG-04: cost-aware rate objective + honest cost accounting (mean_factors/active_R)

Upload src/trident/decoder.py with huggingface_hub

Upload src/trident/byte_lane.py with huggingface_hub

Upload src/trident/layers.py with huggingface_hub

baseline: matched byte-transformer

core: bound retention to [rho_min,1] (kills U/A blow-up)

config: rho_min retention floor (NaN fix)

core: checkpoint passes carried state

data: contiguous byte streams for TBPTT

core: thread state0 through blocks

model: thread recurrent state for streaming eval

recurrence: fp32 autocast guard for chunk scan

config: chunk_size knob (excluded from identity)

recurrence: exact chunked parallel delta-scan (per-head erase gate)

decoder: memory-efficient blocked local attention

Trident 0.2: variable-rate byte architecture, trainer, eval, tests