Drop the refusal patch the claim boundary made redundant

[atomchung]

SKILL.md carried a paragraph reopening a turn that no longer closes.

When consider failed closed on the action, an unanswerable portfolio half stalled the whole turn, so "A refusal does not end the turn…" had to exist to reopen it. #828 moved the limit to the claim — consider fails closed for book-derived claims, "Missing inputs narrow claims, not Skill engagement", "An absent portfolio-derived number stays out" — and the two bullets directly above already state the continue-anyway behaviour the paragraph restated.

So the patch outlived what it patched. Removing it leaves one rule where there were two: a portfolio-derived number comes from the engine; an absent one stays out.

Observed effect this was measured against: given a premise whose ticker the recorded book did not hold, the skill returned "go reconcile first" and stopped in 2 of 3 test premises — the paragraph was present and read in every one of them. A second sentence asking for the same behaviour is not what closes that gap.

• python3 tests/run_all.py --group product — 49/49 pass
• grep -rn "does not end the turn" — no other surface restates it

🤖 Generated with Claude Code

[atomchung]

Read-only review: CLEAN. Verified against main@19250bd4 and the installed-tree references/price-feed.md. The removed paragraph's continuation behavior is already stated by the immediately preceding unpriced/no-book bullets, while the non-degradable forward-consequence boundary remains explicitly owned by price-feed.md (consider refuses rather than answering on cost basis, with the inversion consequence documented). This is a true duplicate deletion, not a safety-boundary deletion. CI is green on head eb379c4; no correctness, privacy, persistence, or compatibility finding. GitHub does not allow the repository owner to approve their own PR, so this is recorded as COMMENT rather than APPROVE.