Teach the policy author and publish skills to write Jev semantic checks

[chhhee10]

Teaches the two policy skills to write and publish Jev (semantic) checks, next to the regex/JS policies they already cover. Everything is checked against failproofai 1.0.8-beta.0 (FailproofAI/failproofai#833): every command and refusal message quoted here was run and captured from that build.

failproofai-policy-author

• New references/jev.md covers:
  • when to use Jev over a regex ("asked for" vs "slipped in");
  • the two tiers (regex is a hard floor, Jev judges above it);
  • what reviewable means, and the rule for marking a policy reviewable: is there anything left that can deny;
  • the check fields, the noul question format, the thresholds, and what Jev is shown;
  • writing probes as claims;
  • a complete pack entry that installs and runs;
  • Jev fields count only in packs (FailproofAI Cloud-managed policies never read them);
  • budget, reserved names, observe/--cli packs, and minCliVersion;
  • local testing, with prompt-wording traps;
  • the 16 built-in checks and 15 reviewable built-ins.
• SKILL.md gets a short "Jev: when no string decides it" section that points to it.
• traps.md §7 was wrong and is corrected: enabledPolicies is ignored once any pack is installed. New §10 lists seven ways Jev fails quietly.
• api.md and cloud.md are updated. scripts/test-policy.mjs now runs regex only, so a configured Jev can't hide what the floor decides.

failproofai-policy-publish

• New "Packs that carry Jev checks" section covers:
  • what publish refuses, and the exact messages;
  • the dry-run output;
  • the question budget (9,101 characters for a pack from outside FailproofAI);
  • the minCliVersion it writes (1.0.8-beta.0);
  • the rollback reminder;
  • what someone installing the pack sees in policies add, policies show and jev status.

Notes

• Needs failproofai 1.0.8-beta.0 or later (#833). Merge this after that release is out.
• One line will follow as its own commit: #833 is tightening when consent may clear a call with several targets, and the skill will state the final rule.

🤖 Generated with Claude Code

[chhhee10]

Superseded by #24. #24 now carries the publish-side Jev content from this PR (the policy-publish section, the refusal messages, the budget, minCliVersion, and what consumers see) and the consent rule from e4756b1. All of it was re-checked against failproofai 1.0.9 and rewritten where 1.0.9 changed it. Closing this one so there's a single PR.