h-compare

# h-compare — Fair comparison with Pareto front You are running the FPF compare workflow: characterize dimensions → declare parity plan → declare selection policy BEFORE scoring → dim-wise parallel scoring → compute non-dominated set → present Pareto front (NOT a scalar winner). ## Step 1 — Ensure portfolio exists If no `portfolio_ref` is given: - Look up via `mcp__haft__haft_query(action="status")` for active portfolios - If only one active portfolio matches the problem, ask the operator to confirm The kernel auto-detects when only one active portfolio exists, but explicit reference is safer. ## Step 2 — Characterize dimensions (agent drafts, operator reviews) If the portfolio's problem has no dimensions, **the agent drafts them** based on the variants and the problem signal. Do not delegate dimension authorship back to the operator — that defeats the value of the agent. Read the variants, identify the axes on which they actually differ, draft 2-5 dimensions inline, then call the kernel. Surface the drafted dimensions to the operator with "Drafted these dimensions — edit if any are wrong before scoring." Operator review is the gate, not operator authorship. ``` mcp__haft__haft_problem( action="characterize", problem_ref="<prob-...>", dimensions=[ { "name": "latency_p95", "role": "target", // constraint | target | observation "polarity": "lower_better", "scale_type": "ratio", "unit": "ms", "how_to_measure": "<single...