verify_task_body cross-issue-reuse-pins check SKIPs on pre-merge worktree bodies — its only gate

Goal

Make verify_task_body.py's "cross-issue reuse pins declared" check actually evaluate on a pre-merge worktree body, so the clean-result footer's reuse-provenance surface is mechanically checked at the one gate that reviews it.

The gap

Surfaced by clean-result-critic while gating #2223 (epm:clean-result-critique, 2026-08-14T09:54:18Z). Its own words:

the cross-issue reuse pins declared verifier check skipped with "eval root unresolved" — but pre-merge worktree bodies are the only place this critic runs, so that skip leaves the footer reuse-bullet surface mechanically unchecked at its only gate.

The structural problem: the check resolves an "eval root" that does not resolve for a body still on its issue-<N> branch, and a kind: experiment clean-result is ALWAYS reviewed pre-merge (the worktree rebase-merges to main only at the terminal step, after the critic gate). So the check is SKIP-by-construction in the only situation it is asked to run. A SKIP reads as "nothing to see" rather than "not evaluated", which is the silent-pass shape.

This was not hypothetical on #2223: the critic had to find the missing axis provenance BY HAND as a Lens 5 blocker — both legs' load-bearing instrument (the #2203 in-house layer-14 axis staged from HF, and Lu et al.'s published Qwen3-32B axis) had no fetchable pin in the body. A working mechanical check is exactly what would have caught that without depending on a reviewer noticing.

Two candidate fixes (the critic named both; pick after reading the check)

1. Resolve the eval root from the ISSUE BRANCH rather than assuming a merged main layout, so the check evaluates in the worktree case.
2. Failing that, DEGRADE to the eval-root-free half of the check rather than skipping outright: assert a Reused:-bullet is PRESENT whenever the body's Goal/Context cites a parent artifact. Weaker, but it fires.

Preference is (1) if the eval root is derivable from the branch; (2) is the fallback. Either way the outcome must be that the check REPORTS rather than SKIPs on a pre-merge worktree body.

Acceptance

• The check evaluates (PASS or FAIL, not SKIP) on a kind: experiment body sitting on its issue-<N> branch pre-merge.
• A body citing a parent-issue artifact with NO reuse-provenance bullet FAILs (or WARNs, matching the check's existing severity class — do not silently escalate a WARN to a FAIL and wedge the fleet Step-9c gate).
• A regression test pins the pre-merge-worktree case specifically, since that is the case that regressed to SKIP.
• Confirm the change does not flip existing green bodies red: run the check across current has_clean_result=true bodies and report the delta before landing.

Provenance

Surfaced as a prose workflow-fix follow-up by clean-result-critic during the #2223 clean-result gate; filed by the orchestrator per the workflow-fix-on-bug protocol (surfaced-prose suggestions get the same auto-file treatment as a formal workflow-fix-candidate block). Not diagnosed further by the filer — the check's code was not read before filing, so the two candidate fixes above are the critic's framing, not a completed diagnosis.