Add 4 verify_plan.py mechanical-backstop checks surfaced during #2502 planning (new-model-env-row, selection-inherited-CI, cross-section verdict-drift, concurrent-launch name-collision)

These are workflow enhancements, not bug fixes: the /adversarial-planner critic ensemble CAUGHT all four classes on #2502 (they became Must-Fixes MF-A / MF-H and standing lens catches). Each proposed check is a cheap mechanical EARLY-WARNING backstop for a class currently caught only by an LM lens or a human critic. All four are heuristic surfaces — the binding gate stays the lens/critic.

Calibration contract (BINDING for every new regex-based check — the c39/c33/c57 precedent): any check added here MUST run the persisted-plan corpus sweep (tasks/*/*/plans/v*.md) at implementation time and record, in a code comment on the check, the realized WARN/FAIL count + distinct-task count + every false-positive eliminated with its named lever. Default posture WARN-only (a FAIL-posture check mis-calibrated on the corpus is the #1388 fleet-wedge shape — the no-flags lint IS the Step 9c gate). Provide the canonical N/A — <reason> standalone-line escape per the existing check convention, and add tests/test_verify_plan.py fixtures (a true-positive plan + a negative-control plan) for each.

Item 1 — new-model-env-row check

Gap (MF-A): a plan introducing a model family the repo-standard stack cannot load (repo-standard vLLM 0.11.0 / transformers 4.57.6 load Qwen2.5-* but NOT Qwen3.5) that does NOT declare a dedicated env/venv provisioning (--env <name> selector + a venv-pins artifact reference) would crash at model-load after a full pod provision. On #2502 the critics caught it (MF-A → the #2378 pod venv port). Proposed check: WARN when a plan names a non-repo-standard model identifier (a model string outside a small allowlist of repo-loadable families) with no co-located --env/venv-pin declaration. N/A escape: N/A — repo-standard model only.

Item 2 — selection-inherited-CI check

Gap: .claude/rules/selection-symmetric-nulls.md requires a bootstrap/resampling CI reported at a max/argmax/best-of-selected axis position to be the SELECTION-INHERITED CI (or both frozen + inherited, labeled), never frozen-at-winner alone. Currently caught only by the c11 statistics lens / statistics-critic (#1434 was caught only at interpretation-critique). Proposed check: WARN when a plan reports a bootstrap/resampling CI (bootstrap/resampl/CI vocabulary) co-located with a max/argmax/best-of axis selection and does NOT name the selection-inherited form. N/A escape: N/A — no CI at a selected axis position.

Item 3 — cross-section verdict-drift check

Gap: c20 (verdict-lattice coherence) checks within the hypothesis/verdict section; it does not reconcile the §3 hypothesis truth-table verdicts against the §7 decision-gate thresholds. A verdict lattice whose Inconclusive/PASS/FAIL cells do not map to reachable §7 gate outcomes drifts silently across sections (MF-C on #2502 required assembling ONE coherent H3 truth table with a reachable Inconclusive across §3/§6/§7). Proposed check: WARN when the hypothesis-section verdict labels and the decision-gate-section outcome set are not in 1:1 correspondence (each declared verdict reachable by some gate outcome; each gate outcome mapping to a declared verdict). N/A escape: N/A — no registered verdict lattice.

Item 4 — concurrent-launch name-collision check

Gap (MF-H): a plan embedding ≥2 dispatch_issue.py launch commands for the SAME issue without distinct --name-suffix values collides on the pod-<N> name (two pods, one name). On #2502 the critics caught it (MF-H → distinct --name-suffix model-a/model-b). Proposed check: WARN when a plan embeds ≥2 dispatch_issue.py launch lines for the same --issue <N> and their --name-suffix values are not all distinct (a bare --issue <N> with no suffix counts as the primary pod-<N> name). N/A escape: N/A — single launch / N/A — distinct name-suffixes.

Provenance: all four surfaced during task #2502 adversarial-planner rounds (2026-08-23). No bug was hit on #2502 — the critic ensemble caught every class; these are the mechanical early-warning backstops.