ablate: MODE=refiner (baseline vs Trident+refiner + test-time budget sweep)

ablate: MODE=operators capacity sweep + KG-05/KG-08 scorecard

recall fix: answer one patch after query (1-patch read latency); verified read exposure

KG-04 runner: dense recall packing (write-capacity stress) + factor_emb/local_attn env

Upload scripts/cpu_bench.py with huggingface_hub

Upload scripts/eval_headtohead.py with huggingface_hub

Upload scripts/cpu_bench.py with huggingface_hub

eval: CPU inference benchmark

eval: head-to-head Trident vs transformer

baseline: matched training script

trainer: parse rho_min env

trainer: skip non-finite updates

eval: CKPT_DIR + result name

eval: CKPT_DIR support

trainer: configurable CKPT_DIR

trainer: stateful truncated-BPTT path

eval: long-context bounded-state experiment

trainer: expose chunk_size env

trainer: label custom profiles

trainer: compile per-block scan, windowed throughput

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