# operator: parameter perturbation
for tuned_value in [retry_threshold, timeout_ms, weight]:
    run suite against UNMUTATED tuned_value
    expect: suite PASSES (this is the baseline)
    # nudge off-target, small epsilon
    mutate: tuned_value -> tuned_value * (1 + epsilon)
    run suite again
    if a BEHAVIORAL test downstream of tuned_value flips PASS -> FAIL:
        # its outcome depends on the value; it exercises behavior the
        # threshold governs
        SUCCESS SIGNAL: a real assumption was pinned and caught
    if only an unrelated flaky/timing test fails:
        NOISE: discard, it means nothing
