zairepo — conv-scale photo-pack lab trusted repo

Full-state snapshot of the v10x4 campaign (budget x8, 25.5K params, 448² rc14 decoder) plus the complete recipe + data pack, ready for a GPU continuation (Kaggle/Colab).

	held-out @224	native 448
v10x4 (this repo)	25.2 dB / 96% NCC / HF 0.63	22.8 dB / 94%
v9x4 (prior best)	22.6 dB / 93% / HF 0.45	20.8 dB / 90%

• state/ — final .bin weights, layout, and full checkpoint (epoch 192, step 79,104)
• recipe/ — the exact trainer + orchestrator + held-out eval, gpu_config_x16.json, and kaggle_x16.ipynb — THE notebook to import into Kaggle (build id: notebook_build_id.txt; build-kaggle-nb.py in the same folder is only the generator — do not import it)
• recipe/gate/ — G1 fixture bundle (replay inputs + TS reference) so the notebook re-certifies step parity on any instance
• data/ — the 200-photo training pack (pack_hash in manifest), zebra/wheel probes, 6 held-out photos
• results/ — held-out val table/figure, full model registry
• manifest.json — sha256 of every file + provenance + the exactness contract (G0–G3 gates)

Exactness — PROVEN, not assumed: the PyTorch port passed all gates on the source box: G0 RNG byte-identical (10k draws), G0b vectorized==scalar, G1 step parity rel 5.4e-08 (8 lockstep steps), G2 eval parity 2.65e-06, G3 state roundtrip sha256 bit-exact. recipe/kaggle_x16.ipynb re-runs the gates on any GPU instance before training.