Phase-3a dual-run: baseline + floor=all-o_proj on granite-8B

Phase-1 sensitivity scores: granite-3.3-8b, 280 layers, n_samples=32 (uploaded from local — worker token can't write to datasets)

v3: pin torch<2.7 to avoid CUDA-13 driver mismatch on worker

v2: emit gzip+b64 JSON to stdout to sidestep worker-token 401

add UV-script runner for granite-3.3-8b sensitivity profiling

initial commit