Strengthen ACS-Core conformance tests with explicit contradictions

Self-audit caught six tests that were positive-only (asserted the
happy path but had no falsifier). A no-op validator or a non-chaining
chain would have passed several of them. Mutation-tested the
strengthened versions: each named bug is now caught with a clear
failure message citing the spec.

Strengthening per group:

  Core02 Envelope shape
    + test_contradiction_validator_actually_works — a deliberately
      broken envelope ({"jsonrpc": "2.0"} with no method/id/params)
      MUST be rejected. Without this, a no-op validator passes every
      drop-required-field test silently.

  Core03 Hook taxonomy minimum
    + Reframed each of 6 hook tests as parametrized over a table of
      (method, valid_payload, broken_payload, schema_file). Asserts
      both that the valid case produces a KNOWN disposition (not
      garbage) AND that the broken_payload is rejected by the
      canonical hook payload schema. Falsifier-per-hook.

  Core04 Dispositions
    + test_allow_response_without_required_envelope_fields_rejected —
      synthesizes 5 broken allow responses (missing type, acs_version,
      request_id, decision; bogus decision enum) and asserts each
      fails response-envelope.json. Without this, the positive
      "allow validates" test is tautological.

  Core05 SessionContext + chain
    * test_chain_externally_recomputable extended to THREE entries.
      Old version only checked entry 1, which doesn't have a
      previous_hash to fold in — a "chain that doesn't chain"
      mutation produced the right value for the root and passed.
      Now: entry 2's recomputed hash with previous_hash MUST match
      what the Guardian published; AND must NOT match the computation
      that ignores previous_hash. Entry 3 verifies transitive chaining.

  Core08 Decision honoring
    * Replaced wire-level "Guardian responds fast" check (wrong
      property) with three adapter-side tests of the actual MUST:
        - test_adapter_actually_applies_guardian_deny — Guardian
          returns DENY, adapter MUST translate to deny, not allow.
        - test_adapter_waits_for_a_slow_guardian — Guardian sleeps
          1s, adapter MUST take at least 1s; an adapter that proceeds
          without waiting is caught by elapsed-time check.
        - test_fail_open_emits_audit_event — unchanged.

  Core10 Wrapped MCP
    * Strengthened the "no crash" test to also require the response
      validates against response-envelope.json. A no-op Guardian
      returning {} would no longer pass.
    + test_mcp_method_namespace_rejects_garbage_namespaces —
      contradiction: methods outside the reserved namespaces (e.g.,
      "arbitrary/method", "step/typo", "PROTOCOLS/upper") MUST be
      rejected by the schema. Without this, "namespace pattern works"
      is unverified.

Mutation-tested the strengthened suite. Four representative bugs were
injected one at a time; each was caught:

  No-op _validate_request_envelope          → 9 Core02 tests fail
  compute_entry_hash ignores previous_hash  → Core05 3-entry test fails at entry 2
  Adapter silently fails open (no audit)    → Core08 audit test fails
  Adapter proceeds without waiting          → Core08 deny + slow-guardian tests fail

Net: still 44 tests, but every one now has a verifiable falsifier.
Adopter who forks and breaks a Core MUST gets a precise failure with
spec citation, not a passing suite that papered over their bug.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: bar-capsule