Differentiate personas by COLOR: weight overall by axis relevance (not unweighted mean)

From P3+ A2/A3 live verification (commit a902c3b). The two seed kitchen personas (value-satisficer, performance-maximizer) produce near-identical COLOR ratings — only the justification prose differs. Causes: (1) the Sonnet orientation pass returns mostly persona-agnostic directions (price=lower_better, wattage=higher_better for both); (2) the overall color is an unweighted mean of dimension colors.

Fix to make maximizer vs satisficer actually diverge on color:

• Weight the overall by each persona's axis relevance (axis_weights in domains/<d>/personas.json), mapping dimensions → axes so a satisficer leans on price_quality/brand_trust and a maximizer on performance dims.
• Optionally let orientation encode per-persona dimension IMPORTANCE (not just direction), so low-importance dims compress toward yellow.

Touch points: services/agents/src/synthesizer/rubric.ts (assembleRubric / overall) + src/rater/rate.ts (rateFacts overall aggregation).