feat(governance): Policy v2 Tier A — set-membership + cross-arg conditions (v0.284.0)

[vikasprogrammer]

First concrete step of the Policy v2 wedge in docs/agent-os-plan.md §4.3. Small, pure, self-contained — no new plane.

What

Extends the JSON rule engine (src/governance/policy.ts) with two new when shapes, still stateless and still first-match:

• in / nin — set membership against an array value → the invalid-enum guard (status ∉ {paid,shipped,refunded} → never).
• argRef — compare one arg to another arg instead of a constant → the wrong-recipient guard (payee ≠ buyer → ask owner).

Why the sweep work is the real change (not evalWhen)

applyProposal is tighten-only, proven by an exhaustive monotonicity sweep (firstLoosening → sampleArgDomains). New ops the sampler doesn't enumerate would silently make that proof unsound. So:

• sampleArgDomains now emits every enum member for in/nin, and seeds both sides of a cross-arg pair with one shared domain so the cartesian sweep hits arg == argRef and arg ≠ argRef.
• firstLoosening's rare huge-product fallback gains explicit joint coverage of each argRef pair — independent per-arg variation can't observe arg == argRef.

Touch-points

All in policy.ts: Op, OPS, new When interface, evalWhen, describeMatch, OP_PHRASE, validatePolicyDocument, sampleArgDomains, argRefPairs, firstLoosening.

Tests

• New scripts/tier-a-policy-test.cjs — 18/18: runtime semantics (in/nin + argRef eq/ne/gt, incl. fail-closed on missing args), validation of both new shapes, and that the monotonicity sweep reaches the new-op boundaries (a real new-op tightening is seen by describeProposal and accepted by applyProposal, never falsely flagged as loosening). Wired into npm run test:governance.
• Existing conformance: 130/130 unchanged — backward-compatible, the bundled default policy uses no new ops.
• npm run build clean.

🤖 Generated with Claude Code