verify_report.py: scan the detailed companion writeup, not just the report body

Goal

verify_report.py's interpretivity / banned-lexicon scan and its structural checks run on the report BODY only. The v2 report template (.claude/skills/issue-v2/report-template.md § The detailed companion writeup) states that the companion at docs/reports/issue_<N>_detailed.md is 100% agent-written and that "the interpretivity rule applies throughout it, and the banned-lexicon scope mirrors the body's". Nothing mechanically enforces that.

Add a generation-mode check that resolves the body's **Detailed writeup:** pin, materializes the companion at that SHA (git show <sha>:<path>), and:

1. runs the same banned-lexicon scan over it — with the Motivation copy keeping the body's hypothesis-framing exemption, since the companion's Motivation is a verbatim copy of the body's; and
2. asserts the forbidden headings are absent — no ## TLDR, no **Takeaways**, no ## Conclusion and next steps. The companion carries no Thomas slots at all, because it is regenerated wholesale on every follow-up round, so anything hand-written there would be silently destroyed.

Why it matters

The companion is the full-detail layer: every figure view, the unabridged methodology. It is longer than the body and it is the document a reader goes to for detail. An asserted conclusion that lands there is exactly as misleading as one in the body, and right now nothing catches it — the whole point of v2 is that no agent asserts a conclusion.

The structural half matters for a different reason: a **Takeaways** block in the companion would read as a claim slot, invite someone to fill it, and then be wiped by the next regeneration.

Evidence this is a real gap (#2162, the first v2 report)

The 18-check verify_report.py output has no companion-content check of any kind. The report-verifier agent scanned the companion BY HAND in round 1 and found 3 lexicon hits — all benign process-facts in the deviations section ("each confirmed from artifacts", "the dispatch log confirmed num_workers=4", "the plan's phrasing implies one fused forward"), none reading a measured value as a finding, so the round still PASSed. But that scan was manual and therefore depends on the reviewing agent choosing to do it.

Worth noting for whoever implements this: those 3 benign hits show a naive scan will produce false positives on legitimate methodology prose. "confirmed" and "implies" have honest process meanings ("confirmed from the artifact", "the plan's phrasing implies"). Consider whether the companion's scan should report hits as WARN rather than FAIL, or whether the lexicon needs process-sense carve-outs. A check that FAILs on "confirmed from artifacts" will be routed around, which is worse than a WARN that gets read.

Scope notes

• Generation mode is the natural home (the pin and the local objects exist by construction there). At promote, a body whose companion pin predates a fresh clone may be unresolvable — degrade to WARN, mirroring the existing image-pin-blob-identity mode-split ladder rather than inventing a new one.
• A grandfathered body with no **Detailed writeup:** line already only WARNs; keep that, and skip this check entirely for those.
• Do NOT weaken any existing check.
• Two sibling verify_report.py tasks are already filed and are NOT superseded by this one: the committed-under + code-SHA asserts, and the stale-evidence-pin check. This is a fourth, distinct gap in the same file — #2162 being the first v2 report, the file is getting its first real exercise.
• Confidence: high that the gap is real and the template already requires the behavior; moderate on the FAIL-vs-WARN posture, which is the one judgement call and is worth deciding deliberately rather than defaulting to FAIL.