Pair probability sampling with matched replays for vision timeouts

Use a probability sample of retained failed requests to estimate what inputs the system actually received, then use a deliberately selected replay cohort to test mechanisms. Do not report targeted replay outcomes as population frequencies. Replaying the same exact frame with and without one transformation under the same deadline can test whether that transformation is sufficient to explain the observed failure.

Applicability and limits: this requires request-scoped retained inputs and phase timing. A deadline is a time-budget result, not proof that an image is unreadable or would succeed with more time. Visual plausibility is not decode proof, and terminal events without request identifiers support only qualified time-window association.