verify_plan c67: scope the temperature-negation to the clause, not the whole line

Goal

scripts/verify_plan.py check c67_retest_kappa_temp0 (landed in #2204) WARNs when a plan registers a test-retest κ demotion gate while pinning temperature 0 in the same judged-DV section. Its negation arm is line-global: a temperature-0 pin is suppressed when ANY nonzero / API-default temperature token appears anywhere on the same line. Decide and implement a narrower scoping (clause- or assignment-scoped) that keeps the load-bearing exception while closing the mixed-stage false negative — and state explicitly, with a measured corpus count, which new false positive the narrowing creates and which residual false negative it leaves open.

Why (measured boundary, not a hypothesis)

Measured on the landed check (kind: experiment, one innermost H3 carrying both the judge line and the retest-κ gate line):

judge line	c67
temperature 0, rubric-keyed cache.	WARN
temperature 0 (not the API default)	WARN
temperature 0, not temperature = API default (1.0).	PASS
Judge at temperature 0; generation at temperature 1.0.	PASS
Judge at temperature 0, falling back to temperature = API default.	PASS

So the negation is temperature-token-anchored: a denial silences the pin only when it carries a SECOND temperature token on the same line. Rows 3–5 are genuine positives that return PASS — the check's own plan discloses them as FN class 5 ("mixed-stage same-line negation"). Row 2 is the shape the #2204 round-1 Codex twin reported as a false negative; the measured behavior says otherwise and the reconciler resolved it against the report. All five rows are now pinned by tests/test_verify_plan.py::test_c67_line_global_negation_boundary_pinned, so this task is a deliberate, test-visible boundary move — not a silent regex tweak.

The cheap fix is not a strict win, which is why #2204 deferred it rather than shipping it. Order-scoping the negation (only a nonzero token BEFORE the zero suppresses it) flips the load-bearing corrected-prose shape into a false positive: #2202 plan v2 L136 and L501 — the prose that FIXED the founding incident — read as "temperature = API default (1.0) … at temperature 0 a byte-identical retest returns near-identical output, κ≈1 for every mode, and the demotion gate could never fire", i.e. the real pin FIRST and the trap QUOTED after. A check that fires on the plans that already fixed the bug is worse than one that misses a compact multi-stage line, so the narrowing needs a scoping rule that separates assignment from explanatory quotation, not just a positional one.

Acceptance

• A scoping rule (clause / sentence / assignment-expression, whichever the plan argues) replaces the line-global not _C67_TEMP_NONZERO_RE.search(line), with the argument for the chosen grain written down.
• The four PASS/WARN rows above are re-pinned to their POST-change intended values, and each changed row carries a one-line rationale in the test. test_c67_line_global_negation_boundary_pinned is updated, never deleted — moving the boundary must stay visible in the diff.
• #2202 v1 → WARN and #2202 v2 → PASS still hold as live CLI runs against the real plan files (the two binding real-fixture criteria from #2204).
• Corpus-calibrated FP/FN accounting: run the new predicate over the retest-bearing plan-version corpus (#2204 measured 90 retest-bearing versions, 63 kind: experiment / 12 infra) and report, as counts, how many versions change verdict and how many of those are correct. A narrowing that raises the armed-kind WARN count without a named true positive behind each new WARN is not shippable.
• WARN-only, never FAIL; the _standalone_na_declared escape, the kind: experiment|analysis gate, and the fence mask are all preserved (each is separately pinned).
• Residual FNs stay DISCLOSED in the check docstring — the markdown-table pin and split-line retest/κ declaration classes are out of scope here and must not be silently dropped from the disclosure list.

Provenance

Deferred from #2204 round 1 by the binding reconciler verdict (epm:review-reconcile v1, 2026-08-19T22:17:12Z, recommendation 1) after a PASS (Claude code-reviewer) vs FAIL (Codex twin) split. The Codex BLOCKER c67-line-global-nonzero-suppression is recorded in #2204's concerns.jsonl with deferred_by: reconciler; this task is where it lands. Filed by the #2204 orchestrator per .claude/rules/workflow-fix-on-bug.md and #2204 plan v2 L353.

Reference points: scripts/verify_plan.py (check_retest_kappa_temp0, _C67_TEMP0_RE, _C67_TEMP_NONZERO_RE), tests/test_verify_plan.py::test_c67_* (11 tests), tasks/*/2204/plans/v2.md § FP/FN surface (FN classes 1–7) + §11 rejected alternatives, tasks/*/2202/plans/v1.md L136/L140 (the founding incident) and v2.md L136/L501 (the corrected prose this must not flag).