README: root cause found and documented

fix root cause: dynamo cache_size_limit 1024 -> 8 (guard-chain scan)

cloud_training README (readable without unzipping)

cloud_training: 4096-ctx notebook, dataset, reqs, README

DPO-on-base adapter (r=16, 1 epoch, 63 steps) - zipped

DPO-on-base bundle: notebook, 252/34 pairs, docs

DPO stage 2: gold DPO adapter + SFT adapter + scripts/logs (corrected reference model)

distilled 9B base (merged 16-bit, alpha16)

repo card

initial commit