Re-encode the 48 BF16 attention output projections of the trial-117 bake into the base checkpoint's MXFP8 format (E4M3 + E8M0 per 32) with constrained rounding that preserves each layer's removed refusal direction (error along it <= 0.002% of the intervention; relative weight error 1.47-1.93%). Changes tensors/model-00001-of-00021.safetensors, config.json, the index, provenance, manifests and the card; adds reencode/. The BF16 bake remains at tag trial117-baked. One-Spark checks: refusals 0/100 dev, 6/100 holdout (BF16 bake 1, 4-6); first-token KL to the bake 0.005; decode 67.8 vs 71.2 ms/step with serve-spark.sh.
