verify_task_body v4: mechanical FAIL for banned Confidence: sentences (check 6 gap)

Goal

Give the SPEC.md rule "confidence lives in the H1 title tag ONLY — no Confidence: ... sentence anywhere in a v4 body" a mechanical arm in scripts/verify_task_body.py. Check 6 currently only asserts the title TAG is present, so a v4 body that also carries a Confidence: ... bullet/sentence PASSes the verifier and the rule is enforced only by reviewer prose.

Provenance

Surfaced in prose by the #2479 interpretation-critic round-2 report (2026-08-24): the round-2 body carried a Confidence:-led ## Takeaways bullet that verify_task_body.py --issue 2479 PASSed; the critic had to flag it manually.

Design sketch

Add a v4-scoped check (new id or an extension of check 6) that FAILs when a line matching roughly ^\s*(?:[-*]\s+)?(?:\*\*)?Confidence: appears in a <!-- clean-result-v4 --> body, outside verbatim-quotation surfaces (fenced code blocks, <details> blocks, blockquoted sample lines — reuse the existing elision helpers if present). Forward-only: v3/v2/pre-sentinel bodies are never newly hard-FAILed (grandfathering contract).

Acceptance criteria

1. A v4 body with a - Confidence: HIGH ... Takeaways bullet FAILs with a named check id; the same text inside a fenced block or blockquote does not trip it.
2. A v3 body with a legacy Confidence: line still PASSes (forward-only).
3. Tests land in tests/test_verify_task_body.py covering both cases.
4. The SPEC.md mechanical-checks section and any check-count references stay consistent.