A must-reject lockfile blocks the merge when a frozen bad input starts returning 200

Handling one lockfile row means importing a target and calling it. `load_target` splits the dotted path on the last dot, imports the module and returns the named attribute ^[14]. `expect_exception` then calls `fn(**payload)`, and if that call returns at all, the detail is "forbidden success: call returned" ^[15]. The fixture has to be a mapping of keyword argument names for that splat to work, and the example fixture is flat: `{"email": "not-an-email", "tenant": "t-1"}`, written with sorted keys and hashed with `hashlib.sha256` ^[19][27].

The class check compares strings. The printed line is `type(exc).__name__ != class_name` ^[16]. The code field narrows it, read off the exception as `getattr(exc, "code", None)` and compared only when the lockfile row declares a code ^[17]. Both example rows declare one: `email.invalid` on the `ValidationError` entry for `billing.api.create_user`, and `invoice.illegal_transition` on the 409 entry for `billing.api.transition_invoice` ^[8][9].

Per case, the work is one JSON fixture, one hash, and one lockfile row added by hand ^[10]. Each row in the example holds eight leaf values, and only `input_sha256` comes out of a tool, which leaves seven set by a person ^[23]. The article's placeholder digests are `c0ffee01` and `bada5502`, eight hex characters where a SHA-256 digest is sixty-four, which is presumably why the author tells you to replace them with real output from `sha256sum` ^[24][10]. The runner re-hashes each fixture on every run, so an edited fixture fails the gate ^[10].

The call happens in process, against whatever the dotted path imports ^[13][14]. A rejection performed upstream of that function, by gateway schema validation or an authz filter in middleware, never reaches the callable the gate invokes, so the gate does not see it ^[26]. Target selection follows from that. The author's advice is to pick the public functions an agent is allowed to touch and collect inputs production already rejects, naming invalid emails, closed-invoice transitions, cross-tenant ids and oversized payloads ^[20]. His stated reason for expecting gaps there is that existing tests assert success, leaving validators, authz checks and parse errors under-represented ^[2].

The runner listing is labeled "proposal, unexecuted in this article", and the published excerpt ends at the `expect_http` signature ^[22]. For this to transfer to another repo, the rejections that matter have to sit behind a function the runner can import, and the enumeration has to keep happening after the week someone set it up ^[20].