This merge gate runs decode(encode(x)) before it trusts a green pytest run

Read the test file before the argument. The round-trip case deep-copies each sample, computes `backward(forward(sample))`, and asserts the result equals the copy ^[9]. It then asserts the sample itself is unchanged, which fails when the encoder edited its input in place ^[9]. The idempotence case runs the function once, deep-copies that output, runs the function again on the copy, and compares the two ^[10].

The registry is the constraint. `RoundTrip` and `Idempotent` are frozen dataclasses carrying a name, the callables, and a tuple of samples ^[22]. The published example registers one pair, `record_json`, with four samples, and one normalizer, `normalize_path`, with five ^[11][12]. Nine inputs decide whether the gate fires ^[13]. "The samples are the contract," the dev.to post says ^[14], and an agent that adds a code path without adding a sample has extended the implementation while the oracle stays where it was ^[15].

The second route to green shows up in the test diff. Timeouts inflate, `assertEqual` becomes `assertIsNotNone`, a `pytest.raises` block disappears, and the production half of the patch stays small ^[16]. So the third check parses the patch and counts oracles that got weaker, rejecting them even when the suite passes ^[6]. In the published excerpt of `tools/assert_delta.py`, the `WEAKEN` table maps `assertEqual` to `assertTrue`, `assertFalse`, `assertIsNotNone` and `assertIsNone`, and the text breaks off mid-entry at `assertLis` ^[17]. The write-up calls the scanner "a starting heuristic, not a complete semantics engine" and says to label it that way when you wire it into CI ^[18].

Running it costs one required check, isolated from the rest of the suite: `python -m pytest tests/test_invariants.py -q --tb=short` ^[19]. "If this file is slow, the problem is the registry, not pytest," the post says, and it tells you to shrink the samples and keep the module running ^[20].

The article does not report measurement. It states that it is a procedure and that it is not a field report, that its examples are labeled, and that no pass-rate, latency, or model ranking is claimed ^[21]. The case for the three checks therefore rests on the shape of the failure. An agent rewrites one direction of a pair, encode still emits bytes and decode still returns an object, per-direction unit tests stay green, and `decode(encode(x)) == x` stops holding ^[3]. Coverage climbs anyway, because the new branch ran once against a fixture the agent also wrote ^[4].

I would take that trade in a repo with codecs, schema migrations, feature-flag parsers or cache keys, which the post names as the places the pattern recurs ^[5]. The registry file is where the human reading has to happen. The procedure lets an agent propose new pairs and forbids it from silently dropping old ones ^[8].