[vyncint]

What is wrong

RunKind::Skipped became reachable in #1112, when has_skip_marker learned the second spelling examples use to decline. Nothing consumes it distinctly. Both places that read the baseline test for Pass and treat everything else alike:

// campaign.rs
if !matches!(baseline.kind, RunKind::Pass) {
    ... println!("schedule-fuzz: BASELINE FAILURE: {:?}; ...")

// cargo-oxide/src/main.rs
Ok(summary) if !matches!(summary.baseline.kind, RunKind::Pass) => {
    eprintln!("Error: baseline example did not pass; ...");
    std::process::exit(1);
}

So cargo oxide fuzz-schedule redux_f32 on anything below sm_100a prints BASELINE FAILURE: Skipped, then Error: baseline example did not pass, and exits 1.

The example did not fail. It ran, declared it cannot run on this device, and exited 0 — which is precisely what scripts/smoketest.sh reports as PASS (skipped). #665 exists because reporting a decline as anything else made an arch-gated example indistinguishable from a broken one, and this is the same conflation in the campaign: sweeping a fleet, "below this example's arch floor" comes back red alongside a genuinely wedged device.

The change

Rather than repeat a matches! in two crates, the distinction becomes one typed definition. RunKind::baseline_verdict() returns Usable, Declined or Broken, and the campaign and the CLI both consult it. The match is exhaustive, so a new RunKind cannot join a class by accident — adding one now fails with E0004: non-exhaustive patterns instead of silently reading as broken.

Behaviour:

baseline	before	after
Skipped	BASELINE FAILURE: Skipped, exit 1	BASELINE DECLINED, a Note: on stderr, exit 0
any other non-Pass	BASELINE FAILURE: {kind}, exit 1	unchanged, and the CLI message now names the kind

--fail-on-finding is untouched and remains the switch for failing on real findings. If you would rather a decline carry a distinct nonzero code, that is a one-line change in the new arm — I picked 0 to match verdict_standard, but it is your call.

Verification

Two tests: the three classes pinned exhaustively over all eight RunKind variants, and that every finding kind is a Broken baseline — a mismatch is a finding in a variant, but as the baseline it means there is no ground truth to compare against. The first fails with Skipped moved back into the broken class; adding a ninth variant fails to compile.

16 ptx-schedule tests and 177 cargo-oxide tests pass, fmt clean, clippy --all-targets -- -D warnings clean on both crates, cargo doc --document-private-items clean.

Two files. No open PR touches either. It shares campaign.rs with my #1134 and #1137; the three touch disjoint regions and I verified all six of this batch merge to a byte-identical tree in either order.

Maintainer repair

• Strong failures now beat skip text.
• The CLI policy is explicit: declined exits 0 with zero seeds, broken exits 1, and usable findings obey --fail-on-finding.
• 17 ptx-schedule and 178 cargo-oxide tests pass, with fmt, strict clippy, rustdoc, and diff checks.

[nihalpasham]

• A skipped baseline had its own verdict, but skip text could still hide a nonzero exit or mismatch, and the CLI exit policy was not tested.
• I layered one commit that fixes the shared classifier and pins the CLI behavior: declined runs execute zero seeds and exit 0; broken baselines exit 1; usable findings obey --fail-on-finding.
• All 17 ptx-schedule tests, 178 cargo-oxide tests, fmt, strict clippy, and rustdoc pass. Hosted CI is queued and waiting for explicit fork-run approval.
• This builds on fix(ptx-schedule): a campaign reads one of the two skip spellings as a pass #1112; no issue closes on merge.

[nihalpasham]

• Declined baselines now stay distinct from broken baselines and real findings.
• Failure precedence and CLI exit behavior are pinned by regression and control tests.
• All seven workflows pass on exact head 327e8d64; both commits have valid signatures and DCO trailers.