task.py set-body: Goal-H2 drop guard is now workflow-aware — v2 report bodies auto-exempt (merged PR #2023)

Goal

set-body refuses to write a kind: experiment body that removes a ## Goal H2 present in the prior body (GoalH2DropError, incident #1112), overridable with --allow-goal-drop. That guard is correct for markdown v4 clean-results, where ## Goal is a required section.

It is wrong for workflow: v2 report bodies. The v2 report template has five required H2 sections — Motivation / TLDR / Methodology (shared) / Results / Conclusion and next steps (scripts/verify_report.py REQUIRED_SECTIONS) — and ## Goal is deliberately not among them; the Motivation section carries that context instead. So every v2 report body legitimately drops ## Goal, and the guard fires on every one of them.

Make the guard skip (or invert) for a body whose frontmatter is workflow: v2 — or, more precisely, key it on the body being a report-v1 body rather than a v4 clean-result. The v4 path must be left exactly as strict as it is today.

Why this matters rather than being a harmless extra flag

The override trains the caller to pass --allow-goal-drop reflexively on report posts. Once it is habitual it is also passed on the runs where the drop is the #1112 accident the guard exists to catch, and the guard stops working for the case it was built for. A guard that fires on a legitimate, spec-mandated action every single time is worse than no guard, because it teaches people to silence it.

What actually happened (#2162, first v2 task)

Posting the assembled v2 report body was refused. Verified before overriding:

• verify_report.py REQUIRED_SECTIONS contains no ## Goal.
• The assembled body has exactly the five required H2s and no others.
• The canonical Goal was never at risk: goal: frontmatter was intact at 1,326 characters before and after the write, and set-body preserves frontmatter. The Goal that downstream agents read is the frontmatter field, which set-goal writes — not the body H2.
• The prior body (with its ## Goal) was snapshotted to original-body.md by --snapshot, so nothing was lost.

So the drop was correct-by-spec and the override was safe here. It was still an override on the happy path, which is the defect.

Scope notes

• Do NOT weaken the v4 path. The check should stay byte-identical in behavior for any body that is not a v2 report; ideally add the branch, do not restructure the guard.
• Decide deliberately whether the discriminator is workflow: v2 frontmatter or the presence of the <!-- report-v1 --> sentinel in the NEW body. The sentinel is the more precise signal (it describes the body being written rather than the task's routing), and it also covers a v2 task whose body is something other than a report. Prefer the sentinel, or require both.
• Consider whether the inverse guard is worth adding for the v2 path: refuse a report-v1 body that is MISSING one of the five required sections, so the same call site protects both shapes rather than just going quiet for v2. That is a natural extension but a judgement call — verify_report.py already covers it at the gate, so this may be redundant.
• #2162 is the only v2 task so far, so there is exactly one real call site to test against; its posted body is a good fixture.
• Confidence: high that the current behavior is wrong for v2 and that the fix is small; moderate on the discriminator choice, which is the one real design decision here.