REAX SN92 demand_forecasting specialist (Plumb-4B line)

A further fine-tune of a Plumb-4B-line decision model for the SN92 REAX demand_forecasting track, trained in the validator's fixed prompt (reax-decide-prompt/1) and single-token option-letter readout. Modified by whynotBeTop: additional LoRA fine-tuning (merged, bf16) and a logit-scale calibration folded into the final norm.

Training data: the public splits of decisionmodels/sn92-specialists (CC-BY-4.0; the demand_forecasting dev split was not used for training), synthetic demand-forecasting items from our own rule-based generators, and public typed-decision data under Apache-2.0 (crh225/plumb-decisions). See LICENSE and NOTICE.