Falsifiability: assert_tool_call red-path proof

[AminChirazi]

Prove assert_tool_call can fail. Fixture: a cassette whose tool-call sequence omits or reorders the required call, plus one whose argument at an asserted dotted path diverges.

Definition of done

1. A NEW fixture under tests/falsifiability/fixtures/ that violates the
   property under test.
2. A test invoking the real assertion path end-to-end as a library call
   (CHARTER §2: every operation is a library call). When a process exit code
   must be checked, capture it directly — never through a pipe.
3. Two checks at two layers: the FAIL verdict, and the non-zero exit code.
4. One appended line in docs/how-flowproof-tests-flowproof.md naming the
   assertion, the fixture, and the defect class it guards.

Hard bars

The Adversary will refuse these even with green CI:

• Do not modify any existing test, committed fixture, assertion implementation,
  or anything under scripts/gate/. Committed cassettes are ground truth
  (§2) — craft new fixtures; never edit a committed cassette into a violation.
• If your red-path test PASSES where you expected FAIL, you have found a real
  false-green defect. That is a RESULT, not a task. Report it with the fixture
  attached for the Ledger keeper; the fix is a separate issue you must not
  self-authorise.
• Fixtures are deterministic and offline. Replay makes zero LLM calls and
  requires no API key.

Charter: Milestone 2 criteria 6–8 (amendment 2026-07-30). Brief:
docs/loop/milestones/m2-falsifiability.md. Parent: #61.

[AminChirazi]

Partly covered — needs narrowing, not a new fixture. a_failing_assertion_refuses_the_trace (agent_flow_e2e.rs:778) is already a red-path proof for assert_tool_call: the flow demands book_flight, the agent never calls it, and record refuses the trace.

What is NOT covered is the argument layer. The matchers (where <path> equals|contains|matches|exists|is absent) have no red path at all — nothing proves that assert_tool_call: create_booking where flight.id equals KQ311 FAILS when the recorded argument is KQ999. Since docs/agent-testing.md calls argument assertions "usually where the bugs are", that is the proof worth writing.

Recommend re-scoping this issue to the argument matchers specifically.