workflow-fix: check-17 verbatim equality for origin_prompt footer quote

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a prose follow-up raised by the clean-result-critic on task #813 (emitting agent: clean-result-critic, round 1 of the per-example-vs-averaged-map follow-up gate).

Goal

Extend verify_task_body.py check 17 to enforce normalized verbatim equality between the **Context:** footer's "Originating prompt, verbatim" blockquote and the frontmatter origin_prompt, FAILing on prefix-truncation.

Workflow gap

• Bug observed: task #813's re-folded clean-result footer carried an "Originating prompt, verbatim" blockquote that was a strict mid-sentence prefix of the frontmatter origin_prompt (cut after "SAVE+UPLOAD the pre and post trained maps,", dropping the unreduced-activations directive) — and check 17 PASSed it; only the LM-side Lens 5 review caught it.
• Why it is a workflow gap: the spec's check 17 verifies the Context row ships the originating prompt, but does not compare the quoted text against origin_prompt, so "verbatim" is unenforced mechanically — any truncation survives to the final gate and burns an ensemble REVISE round.
• Confidence (emitter): high

Proposed change (candidate diff sketch — refine in planning)

In scripts/verify_task_body.py check 17: extract the blockquote under the "Originating prompt, verbatim" label, normalize whitespace on both sides (collapse internal runs, strip), and require equality with normalized frontmatter origin_prompt; FAIL with a context-origin-prompt-mismatch message naming the first divergence offset. Keep a NO-OP pass when frontmatter origin_prompt is absent (legacy tasks). Add regression tests in tests/test_verify_task_body.py (equal → PASS; prefix-truncated → FAIL; absent frontmatter → NO-OP).

Scope / surfaces

• Primary target: scripts/verify_task_body.py
• Also: tests/test_verify_task_body.py, and the SPEC.md check-17 prose if its wording needs the equality rule named.
• Grep the workflow surface for check-17 references before editing (grep -rln 'check 17\|origin_prompt' scripts/verify_task_body.py .claude/skills/clean-results/) and update every hit; list them in the plan.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; SPEC.md stays consistent with the verifier.
• Forward-compatible: v3/v2/legacy bodies are never newly hard-FAILed (gate the new equality FAIL on the v4 sentinel, WARN-only below it).
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).

Provenance

• workflow_fix_target: scripts/verify_task_body.py
• fingerprint: 9dc0f6bb1a35

Surfaced prose (clean-result-critic, task #813 r1): "the fix is a one-line append (mechanizable — a check-17 normalized-equality extension is sketched in the verdict note for the orchestrator to route)". Full verdict: /tmp/issue-813-clean-result-critique-r1.md (epm:clean-result-critique v1 on #813, 2026-07-05).