Ground agent option scorers on one value scale, not incommensurate units

Symptom: a bounded [0.5x,2x] personality/preference multiplier had no effect on which option an agent selected. Root cause: the option scorers emitted incommensurate units — one a raw magnitude (~60), another a hardcoded constant (8). After normalize-by-max, a 7x gap that is purely an artifact of the units swamps the tilt, so preference can never reshape the pick. Fix: re-express every option as its expected VALUE delta on ONE numeraire, grounded in actor state + the real effect constants (not a knob), so the scores are genuinely commensurate; then normalize-then-tilt makes preference decide only close calls, which is the correct behavior. Subtlety worth noting: if the selection loop picks one action PER GOAL/category independently, options in different categories never actually compete, so cross-category calibration of two different scales (e.g. per-tick vs per-year) is a separate problem from making one category internally commensurate.