verify_plan.py: check for inherited argparse row-count defaults under-covering the plan's registered per-cell target n (#2054 --target-conv-ids class)

Provenance

workflow_fix_target: scripts/verify_plan.py Surfaced by: Methodology critic, task #2054 plan v11 round 1 (same-issue follow-up round coordinated-common-set-regen, 2026-08-10). Report: /tmp/issue-2054-critic-methodology-v11-r1.md (session-local; the incident summary below is self-contained).

Gap

verify_plan.py has no check for the inherited-argparse-default under-coverage class: a plan reuses a generation/splice/capture script whose argparse integer default silently caps row/conversation coverage BELOW the plan's own registered per-cell target, and no embedded command overrides the flag. The plan reads as if the target is met; the run truncates deterministically (first-N prefix) and every downstream gate that runs before the consuming phase reads PASS without seeing the breach.

Incident

Task #2054, amendment plan v11 (round coordinated-common-set-regen): the plan registered a ≥2d per-cell training-size target (n_train ≈ 7,200, d=3,584, intersection target |S| = 9,000), but the reused --target-conv-ids argparse default (8,000; phase_c.py:604 / phase_d.py:709 on the issue-2054 branch, deterministic first-N prefix truncation) would have capped 28 of the 48 in-scope cells below the target (n_train 6,400 best case, ~5,700 under divergent per-variant orderings) for a majority of the 336 affected pairs. The round's pre-registered gate 1 runs BEFORE Phase C/D, so it structurally could not see the breach — the same fail-open class the round existed to fix. Caught only by the Methodology critic in round 1; the mechanical pre-pass (verify_plan.py) passed the plan clean twice (v10, v11).

Family: #1727 (smoke-valued variable leaking into production), #1345 (missing-filter class) — inherited defaults silently narrowing a new design's coverage.

Sketch (from the critic report, verbatim class)

A verify_plan-style check: for each §4-named reused generation/splice script, extract argparse integer defaults matching (target|max|limit).*(conv|rows|ids); FAIL/WARN when the plan's stated per-cell target n exceeds a default and the flag token never appears in the plan text.

Implementation notes for the planner (not binding):

• WARN-only to start (the heuristic has false-positive surface: scripts resolved from plan text, branch-resident scripts need git show <branch>:<path> resolution, and "stated per-cell target n" extraction from plan prose is fuzzy).
• Needs a canonical standalone N/A escape phrase (e.g. N/A — no inherited row-count defaults) per the _standalone_na_declared convention, wired into the adversarial-planner SKILL.md escape-phrase list.
• Tests in tests/test_verify_plan.py with fixtures reproducing the #2054 shape (default below target + flag absent → WARN; flag present with explicit value ≥ target → PASS; N/A line → PASS).
• Respect the existing check-id numbering + bundling conventions (workflow_lint.py --check-lessons-index untouched; this is a verify_plan check, not a rule file).

Acceptance criteria

1. New check in scripts/verify_plan.py implementing the sketch (WARN-only acceptable), with check id + name following the existing convention.
2. Canonical N/A escape line recognized standalone-unwrapped, added to the adversarial-planner SKILL.md § canonical escape phrases list.
3. tests/test_verify_plan.py fixtures covering: the #2054 incident shape (WARN), explicit override (PASS), N/A declaration (PASS).
4. Full-suite verify_plan tests green.