Smoke + reform validation: one shared baseline and branched reforms, not 85 standalone simulations

Audit finding 1 of #468, split out as its own increment. Smoke executes 40 full simulations (1 baseline + 39 probes) and reform validation 45 more, each reloading the written H5 and constructing an independent full-release simulation, when only the reform parameter patch varies. Remedy: load the dataset once, one baseline system/simulation, branch or clone per reform (the #461 family-system pattern), cache baseline totals by (measure, period), share content-addressed reform totals between smoke and validation. Verification bar: byte-identical scoring vs independent simulations on a pinned slice (the #461 oracle pattern). Biggest single build-time lever after the #463 loop deletion.

Start by reading the #461 family-system pattern and oracle approach, then trace the smoke and reform validation paths that reload the written H5 and construct full-release simulations. Done means one loaded dataset and baseline are reused, reform totals are shared where applicable, and scoring is byte-identical to independent simulations on the pinned slice.