Make the final self-check agree with the removal patterns

[Scott-Emberson]

The rewrite process's step 3 says "Treat any unsupported addition or lost claim as an error" with no carve-out, but seven patterns require removing claims: §2's After drops two named citations and a follower count, §25's After is "Cut the paragraph", §6's After drops "strategic location and ongoing initiatives", and §5, §21, §34, and §35 all instruct removal. Rule 2 ("Keep every claim.") had the same conflict. An agent applying step 3 literally has to reject the skill's own model answers, so behavior on those patterns becomes inconsistent across runs.

Changes:

1. Step 3 now reads: unsupported additions are always errors; a lost claim is an error unless a pattern above requires removing it (name-drop lists §2, unsourced claims §5, stock outlooks §6, guesses §21, send-offs §25, unraised objections §34, fake alternatives §35).
2. Rule 2 becomes "Keep every supported claim." so the two rules agree.
3. The em-dash false-positive item ("Em dashes are evidence only when paired with formulaic sales-y rhythm") gave the opposite answer to §14 for dash-using input with no writing sample. It now states it governs detection only and that §14 still governs the final rewrite.

This also addresses #212: with the carve-out explicit, patterns that are NOT on the removal list (like §10 rule-of-three and §24 hedging, the ones #212 demonstrates) can no longer read a deleted ranking or simultaneity claim as acceptable shape-editing; step 3 flags it as a lost claim.

SKILL.md stays at 456 lines and the 35 patterns are untouched. No version bump to avoid colliding with #233; happy to fold a release note into whichever lands last.

Closes #212.

🤖 Generated with Claude Code