Extend verify_plan.py with a decision-margin satisfiability probe (margin-vs-baseline ceiling)

Provenance

workflow_fix_target: scripts/verify_plan.py Surfaced by the #2203 full-rerun-bugfix plan-critique (2026-08-10): the Statistics AND Alternatives lenses INDEPENDENTLY caught that plan v12's §3 verdict lattice registered H1/H3-confirm as an ABSOLUTE ≥ 10pp harm reduction while the realized baselines were 7B 9.66% (48/497) and 32B 4.02% (20/498) — so the confirm branches were arithmetically UNSATISFIABLE and a false FALSIFY would ship by construction. The mechanical verifier (c27, and the check-20 verdict-lattice check) did NOT catch it.

Problem

verify_plan.py check c27 covers only SAME-LINE explicit ratios; it has no cross-section check that a registered absolute-reduction decision margin is ACHIEVABLE given the DV's baseline rate. A plan can register "Δ ≥ N pp reduction vs baseline" with a baseline < N% → the confirm branch can never fire and the falsify verdict is predetermined regardless of data. This is the THIRD recurrence of the #810 margin-vs-ceiling family (#810; #825 v17/c27; #2203 v12).

Proposed fix

Extend verify_plan.py (c27, and/or check-20 the verdict-lattice check) with a cross-section satisfiability probe:

• Extract ≥ N pp (absolute percentage-point) reduction-margin clauses from the verdict lattice / §3.
• When the DV's baseline rate is stated in-plan (e.g. "baseline ~9.7%") OR recoverable from a named prior artifact the plan cites, compare: WARN when the registered absolute margin ≥ the recorded baseline rate it subtracts from (the confirm branch is unsatisfiable).
• WARN (not FAIL): baselines are not always in-plan; a WARN with a clear message ("registered absolute margin N pp ≥ cited baseline B% — confirm branch unsatisfiable; use a relative margin or size the absolute margin to the baseline") lets the planner fix it or carry it with a reason.
• Add a pin test (tests/test_verify_plan.py) reproducing the #2203 v12 shape (absolute ≥10pp margin vs a 9.7% / 4.0% in-plan baseline → WARN) and a negative case (relative margin, or absolute margin < baseline → no WARN).

Acceptance

• New/extended check WARNs on the #2203 v12 shape and passes clean on a relative-margin or baseline-sized-absolute-margin plan.
• tests/test_verify_plan.py pin tests added (positive + negative), full workflow-lint + the mapped verify_plan tests green.
• No regression to existing c27 / check-20 behavior.

Notes

Kind: infra (workflow-surface fix; no experiment, no promotable clean-result — completes on the Step 9c test-verdict path). Neutral-vocabulary task; no harmful-content surface.