EBVT instrumentation: expert-usage probe (forward hooks, layer-cake self-check) + baseline X8 103M + harness wiring

docs: re-run verification on HF repo (2026-09-03) + fixed harness

docs: X8 unified checkpoint — sandbox live operation (2026-09-02)

Upload scripts/fast4gpu_boost_v4.py with huggingface_hub

Upload scripts/hf_push_x8.py with huggingface_hub

v3 trainer + English docs + production x8 results

ops: one-shot x8 pod deploy script

fix: atomic checkpoint save (tmp+os.replace)

opt: cumsum/chunked attention kernels, memory-flat CE, block checkpointing, v2 trainer (proven equivalent, 46 tests)

Upload scripts/prep_kuramoto_fix_resume.py with huggingface_hub

Upload scripts/fast4gpu_boost.py with huggingface_hub

Upload scripts/smoke_load_merged.py with huggingface_hub

Upload scripts/fast4gpu_boost.py with huggingface_hub

Upload scripts/fast4gpu_stage2_ss.py with huggingface_hub

Upload scripts/fast4gpu_stage2_fine.py with huggingface_hub

Upload scripts/fast4gpu_surgery.py with huggingface_hub

Add production multi-GPU training scripts (B=2 + compile + shards)

Add production multi-GPU training scripts (B=2 + compile + shards)

Add production multi-GPU training scripts (B=2 + compile + shards)

Upload scripts/fractus_self_dataset.py with huggingface_hub

Upload folder using huggingface_hub

Upload scripts/train_salience_head.py with huggingface_hub

Upload scripts/build_quality_corpus.py with huggingface_hub

Upload scripts/train_1b_gpu.py with huggingface_hub

Upload scripts/train_progressive.py with huggingface_hub