skill-authoring §7: bare arm is environment-relative; harvest the arms' reasoning, not just verdicts

[firaen22]

Summary

Two clauses folded into the existing §7 bare-probe bullet, distilled from a downstream consumer's first four fold probes of this repo's merged rules (pre-registered arms, weaker-tier executor):

• The bare arm is bare of THIS rule, not of the world. It runs in the executor's real baseline — standing global rules, always-loaded caches — since that's the live counterfactual for the fold decision. A bare ✓ therefore means redundant in that environment, not redundant everywhere: one probe's fact-anchoring candidate failed to discriminate purely because the consumer's global rules carry a standing verify-before-relaying order (the general form of the candidate). Record the baseline next to the verdict; never export a non-discrimination result to an environment with a different baseline.
• Read each arm's stated reasoning before writing the fold, not just its pass/fail. A failing bare arm hands you the exact rationalization the folded rule must refute — one bare arm defended shipping an unverified capability doc with "the document didn't lie — it described the intended behavior," which became the folded rule's named rebuttal. A passing ruled arm can still over-fire (satisfying the rule by refusing the surrounding task outright) — one ruled arm refused to produce any plan at all, which is a wording warning, not a win; the fold was reworded so compliance composes with the task.

Both clauses ship unprobed per the covenant; a Provenance paragraph documents the source incidents and queues their own probes.

Test plan

• git diff --word-diff against the branch point — zero words removed from the existing bullet, only inserted
• Line-width and non-ASCII sweep — no new stray characters
• Both clauses extend the existing bare-probe bullet (matching how prior r4–r6 revisions landed) rather than adding sibling bullets
• Provenance paragraph added in house style, anonymized consumer framing

🤖 Generated with Claude Code

[firaen22]

Pushed a third clause to this branch (16e6e1f) rather than opening a follow-up, because it cannot ship separately from clause 1.

Separate baseline from leakage in the bare arm. Clause 1 says the bare arm keeps its environment (standing global rules, always-loaded caches). This clause says strip episodic recall of the finding under test. Stated apart, each reads as contradicting the other — the boundary between them is the actual rule, so they need to land together.

The test is what the arm would have without you: what every invocation in that environment ships with is baseline; an auto-memory store, a session summary, or notes from the run being probed are leakage. That store survives a "fresh invocation", so the bullet's existing no-shared-state requirement does not catch it — the arm loads the conclusion, restates it, and scores as an unaided reproduction.

Countermeasures added: recall-dodging framing (generic scenario, none of the finding's names or phrasing), an explicit disregard-prior-findings instruction, and a citation check on the output before the run counts — an arm citing the finding or its vocabulary is re-run, not scored.

Provenance updated to cover all three clauses; ships unprobed per the covenant.