macroonz-harness

The taste tester. It has no loyalty to the baker, no opinion about the recipe, and it records everything.

You describe a subject once — what it takes, what it returns, what it refuses, what must hold — and its instruments spend the rest of their time trying to make that description false: independently callable engines a caller composes, not one pipeline behind one button. A verdict comes back as two separately earned values that join on one execution key: the report, which carries the standing, the site, and the complete denominator of what ran; and the replay capsule, which a completed reduction mints over the exact minimized input — so a failure is a thing you can hold rather than a thing you saw once, and neither value claims what the other proved.

The harness belongs in test and tooling code, and it never depends on the crates it judges. The macroonz facade includes its API by default so one dependency opens the complete machine; a shipping crate that wants no judge selects the diet posture.

What you write

A row per trial, in your own tests, through the trial_table! stamp or by hand.

macroonz_harness::trial_table! {
    /// Every trial this crate authored for its lots.
    mod lots named("bakery", "lots") {
        provenance: unproduced,
        invocation: InvocationProfile::declared(
            CaseBudget::declared(64),
            ByteBudget::declared(4096),
            TimeBudget::declared(1_000_000_000),
        ),
        target: TargetBinding::bound(
            TargetTriple::declared(TARGET),
            ToolchainIdentity::declared(TOOLCHAIN),
        ),
        clock: HarnessClock::unavailable(),

        suite merge named("bakery", "merge") {
            a_full_lot_refuses_one_more: {
                let (row, attachment) = merge_parts("a-full-lot-refuses-one-more")?;
                Binding::bound(row, attachment, Provenance::Unproduced)
            },
        }
    }
}

merge_parts is yours: an ordinary function composing the Row — its claim, suite, classification, subject, check, population, and origin — and the ExecutableAttachment that runs it, through the descriptor home’s public constructors. Every clause and every row ends with a comma, because the stamp’s grammar says so and refuses the one that doesn’t.

One stamp writes the authored world: an aggregate test per suite and a named lens per row. A row names a claim (what is being established), a subject (the road under test), a check (what decides), and a population (where the inputs come from), each as a namespaced name you own. The harness stores your names, hashes your names, and never reads inside them.

target and clock are declared at the invocation because the harness derives no host fact of its own: a cache key with a guessed toolchain in it is a lie with a digest.

What it does with it

flowchart LR
    accTitle: Harness evidence composition
    accDescr: Descriptor, generation, and scheduled adversity enter the runner, whose report can feed mutation, oracle, and benchmark evidence roads.

    D["descriptor<br/>a row says what it claims"] --> S["runner<br/>select · bind · execute"]
    G["generate<br/>inputs from a seed"] --> S
    F["fault · clock<br/>adversity on schedule"] --> S
    S --> R["report<br/>verdict + evidence · replay minted by reduction"]
    R --> M["muterprater<br/>mutants of the subject"]
    R --> O["oracle<br/>an independent second opinion"]
    R --> B["bench<br/>work under a pinned profile"]

What stays yours

Every seat where meaning enters is a type parameter with no bound, a capture-free function pointer, or a name you declared.

• Equality is your Equivalence; order is your Order; a refusal is read through your ResponseReading.
• A check hands back a TrialConclusion: passed, or refused with a finding whose cause is a (family, local) pair you spell.
• A mutation discovery names points, claims, and operator families you declared; the harness intersects them with your own permission table and invents no attribution.
• The artifact under a structural read is described as plain strings — a target, a trait path, attributes — that you wrote beside the declaration you handed the producer.

There is no trait a subject must implement.

Evidence, not green

A passing lane is a side effect. What the harness is for is the record: which claim was exercised, under which revision of the subject and which revision of the check, on which target, under which budgets — that is the report’s half. The reproduction is the replay capsule’s half, earned separately: a completed reduction mints one over the smallest input reached under the declared reducers and budget, bound to the same execution key the report carries, so the two halves join without either claiming the other.

Four things a verdict never does: misattribute a failure, drop an observation, fabricate evidence, or claim a kill without a qualified baseline.