[vyncint]

What is wrong

has_skip_marker accepts skipping: and nothing else. The convention it is implementing has two spellings, and scripts/smoketest.sh — which owns it — accepts both:

if grep -qiE '^[[:space:]]*(skipping:|pass \(skipped\))' "${log}"; then

Its own comment names the second form in so many words: "PASS (skipped): ... form below sm_75". generated_ldmatrix prints exactly that:

if compute_capability < 75 {
    println!(
        "PASS (skipped): ldmatrix.m8n8.x4.b16 requires sm_75+; device is sm_{major}{minor}"
    );

Why it matters

An example declining that way exits 0 and prints no mismatch marker, so run_binary classifies the run RunKind::Pass. The campaign gates on precisely that:

if !matches!(baseline.kind, RunKind::Pass) {
    ...
    println!("schedule-fuzz: BASELINE FAILURE: {:?}; no schedule variants were run", ...)

With the right kind the campaign stops there and says so. With Pass it proceeds: it patches the PTX, sweeps every seed, and each variant declines identically — same exit code, same output, so not even OutputChanged fires. The campaign reports a clean sweep of a kernel that never launched.

This is #665's bug in the second consumer of the same convention. That fix put it as "a skip the function fails to recognise is not merely missing its (skipped) annotation — it is promoted to a full execution PASS", and taught verdict_standard the second spelling. ptx-schedule's campaign arrived later (#1065) with only the first. A silent all-clear is the worst outcome available to a bug-finding tool, and it is the one this produces on any device below an example's floor.

The change

Hoist both spellings into a SKIP_MARKERS constant and check the line against each, with the coupling to the smoketest's pattern recorded next to it.

Case-insensitivity is unchanged and already correct: run_binary lowercases the combined output before classifying, which is what makes this match the smoketest's grep -i.

Tests

Two:

• both_smoketest_skip_spellings_are_recognised — both forms, with and without leading whitespace; and the near-misses are rejected: a bare pass, pass: 1024 elements verified, success, and a mid-line no skipping: here, since the marker is anchored to the start of the line in the smoketest too
• a_declined_run_is_skipped_not_passed — the two forms as they actually arrive, newline-terminated

Both fail with the second marker removed.

One note, out of scope

smoketest.sh itself is not uniform about this: verdict_standard (line 397) takes both spellings, but the ltoir verdict near line 609 greps only ^[[:space:]]*skipping:. That one is inert today — every LTOIR_EXAMPLES member uses the first spelling, and generated_ldmatrix is standard, so it goes through verdict_standard. I have left it alone since #1102 is currently editing that script; happy to follow up once it lands if you would like the two verdicts aligned.

Verification

All local, no GPU needed.

• cargo test -p ptx-schedule --all-targets: 11 passed, 0 failed
• cargo fmt --check -p ptx-schedule: clean
• cargo clippy -p ptx-schedule --all-targets -- -D warnings: clean
• cargo doc --no-deps --document-private-items -p ptx-schedule: clean

One file, crates/ptx-schedule/src/campaign.rs. Checked against every open PR's file list via gh api --paginate — no overlap, including with my #1108/#1109/#1111 on this crate, which touch Cargo.toml, README.md, main.rs and lib.rs respectively.

[nihalpasham]

• Confirmed both halves: smoketest.sh accepts two skip spellings while has_skip_marker knew one, so generated_ldmatrix's "PASS (skipped)" decline exited 0 with no marker, classified as Pass, passed the baseline gate, and the campaign swept seeds against a kernel that never launched. Both tests fail without the fix; no false positive is possible since a bare "pass"/"pass:" line cannot start with "pass (skipped)".
• The SKIP_MARKERS constant documents the smoketest coupling at the right place; the convention's owner stays smoketest.sh.
• Two pre-existing residuals for a follow-up issue, not this PR: gemm_sol_final and tcgen05_matmul's arch-gated fallbacks print no marker at all, so their declines still read as Pass baselines; and has_skip_marker's case-insensitivity relies on callers pre-lowercasing.