fix(apply): lint reads the same artifact shape apply consumes (A3)

[henleda]

Fixes a false positive in the pre-apply policy linter. Spotted from real scan output:

  triage missing-authz-scenario-003 -> service_policy(full*), malicious_user(partial)
    ⚠ lint deny-unauth-x402-scenario-control: no DENY rule — refine will correct at apply

That policy has a DENY. The linter just couldn't see it.

Cause

generate emits a service-policy artifact in one of two shapes — a bare spec, or a full XC
create body {metadata, spec}. apply_from_scan handles both, and says so:

# Generated artifacts vary — some are full {metadata, spec} objects, some a bare spec.
if isinstance(art.get("spec"), dict) and art["spec"]:
    spec, src_meta = art["spec"], (art.get("metadata") or {})

lint_service_policy didn't. It read spec["rule_list"] straight off the artifact, which on a
wrapped one is None — rule_list sits a level deeper — so it saw zero rules and returned
["no DENY rule"].

In a 12-policy crAPI/x402 run, 6 policies were wrapped and all 6 drew the false positive:

WRAPPED  ['no DENY rule']  deny-negative-pay-amount.json
WRAPPED  ['no DENY rule']  deny-unauth-x402-scenario-control.json
bare     []                deny-sqli-login-trigger.json

deny-unauth-x402-scenario-control contains rules[0].spec.action == "DENY" followed by an
ALLOW — exactly the shape the linter exists to approve. The cost isn't cosmetic: noise at a
50% false-positive rate trains an operator to skim past the one warning that matters.

Fix

Adds artifact_spec() — the same unwrap rule apply_from_scan uses — and applies it in
lint_service_policy, so the two can't drift apart again. Re-linting the artifacts from a real
run: 6 flagged → 0.

Scope

Deliberately not applied to lint_api_schema. Its consumer (console/app.py
_dispatch_action) uploads the artifact to XC's object store verbatim, without unwrapping —
so a wrapped api_schema genuinely would be broken and the linter's complaint there is a true
positive. Every api_schema artifact on disk is bare today; unwrapping in the linter would only
hide a future breakage. Noted rather than papered over.

Verification

• 4 regression tests covering both directions: a wrapped policy with a DENY lints clean, and a
  wrapped policy with no DENY, or with an ALLOW before the DENY, is still caught — the fix
  must not blunt the linter.
• Full suite 190 pass, ruff clean.
• Re-linted every service_policy.*.json artifact on disk: 0 flagged.

Note

Pre-existing, from d751aa1 (the original A3 linter) — not introduced by #1. Whether #1's
switch to a full scan transcript is why it became noticeable is unclear: with ~10 findings the
lint lines may or may not have fallen inside the old 40-line window.

🤖 Generated with Claude Code