Frontier-resolution certificate + near-duplicate option advisory

[cafzal]

What changed

Two additions, both generalized from patterns surveyed in external optimization tooling and reduced to what Frontier didn't already have.

frontier_resolution block in the explore certify certificate (engine/explorer.py). Where coverage/completeness audit the NSGA and exact runs against each other, this bounds the reported frontier against the true continuous frontier: on a 2-objective LP/QP overlay, the certified points' chords (above, by convexity) and their own ε-floor shadow-price tangents (below, by the envelope theorem) sandwich the tradeoff curve; a dual-free monotonicity floor covers segments without a priced tangent, so a missing dual loosens the bound rather than voiding it. The widest chord-to-envelope gap is a certified resolution — "between the anchors, no feasible plan improves the primary over straight-line interpolation of the certified points by more than X" — the frontier-side sibling of explore audit's whole-region guarantee. Consumes only duals already on the run (no extra solves); present-as-null on shapes that can't carry the claim (3+ objectives, integer overlays, no tension), matching completeness. The widest segment is named with its endpoint solution ids as a targeted-tightening handle.

near_duplicate_options advisory in data_metrics (engine/metrics.py). Pairs of complete-scored options within 2% of each objective's score spread on every objective — the "same bet entered twice" case strict dominance structurally misses (an exact tie dominates nothing). Advisory only: capped at 10 pairs + total (the dominated_options convention); the engine flags, the user merges or confirms. A dominated-by-a-hair twin rides both lists; the skill copy says the twin read ("pick either") is the truthful framing there.

Shared cleanups from review: _floor_dual + _FLOOR_ROLES extraction (the ε-floor role tuple had 6 literal sites), _complete_options shared between the dominance and twin detectors.

Docs: architecture.md (certify tool row, Solver Backends paragraph, data-metrics diagram), README (one clause), solution_interpreter › Reading the Certificate (new block paragraph, fixed a stale block count), data_collection › Score Quality Signals (twin bullet), explore tool docstring.

Review and testing

• Full suite green: 894 passed / 2 skipped, including resolver-integrity, wire-compactness, and UI suites.
• New tests: hand-checkable sandwich geometry (parabola with exact duals — gap values verified by hand), monotonicity-floor survival with a missing dual, widest-segment localization, 3-objective / no-duals / no-tension gates, and a real HiGHS end-to-end on an inline 2-objective mean-variance QP (every bundled example is 3+ objectives, where the block correctly returns null). Twin detection: exact ties, near-ties, both-lists case, flat-axis handling, cap+total, incomplete-score exclusion.
• Four-angle quality review (reuse / simplification / efficiency / altitude) applied; skips noted in the plan doc.
• User-tested per the frontier-dev method: in-process handler flow, a real MCP stdio-transport drive (both payloads survive the wire; live QP produced a 0.1%-of-span resolution with all 52 segments dual-tightened), and an agent-in-the-loop session on a realistic five-fund allocation task — the agent surfaced the twin pair as its top verification caveat and presented the resolution claim correctly scoped, without inflating it to global optimality. The live installed server was not restarted (the main checkout is on another session's in-flight branch); the wire test spawns the worktree server via the test_mcp_stdio_source.py bootstrap.

The agent test also surfaced pre-existing payload frictions outside this diff (the exact: false flag beside solver_used: "highs"; knee rationale at corner-adjacent inflections; regret id-space mixing; hypervolume scale mismatch across payloads) — filed separately, not addressed here.