Add a rule against unrequested edits to adjacent code

[Lockyer228]

Why

ponytail: and the repo's own AGENTS.md are explicit that agents must not "improve" adjacent code while fixing a bug — but the skill's Rules never said it. The result is the exact drift the issue names: unrequested edits to neighboring code, comments, or formatting riding along with a real fix.

What

Add a surgical changes rule to the Rules section of AGENTS.md and skills/ponytail/SKILL.md:

• touch only what you were asked to touch
• don't "improve" adjacent code, comments, or formatting; match the style that's there
• remove only what your change orphaned; mention unrelated dead code instead of deleting it

Propagated byte-identically to the seven adapter copies (.cursor, .windsurf, .clinerules, .agents, .qoder, .github/copilot-instructions, .kiro), per check-rule-copies.js, and regenerated the .openclaw/ skill.

Tests

• node scripts/check-rule-copies.js → exit 0 (9 rule invariants intact, no existing rule reworded)
• node --test tests/openclaw-skills.test.js → 18/18
• npm test → all pass except the pre-existing csv: correct pandas one-liner passes failure, which fails identically on the base commit (no pandas installed locally) — unrelated to this change.

Fixes #640