Before the Data Arrives: Staleness Is Not a Flat Line

An unavailable recent measurement is not evidence that a process stayed constant. Yet forward-filled missing blocks are used to evaluate forecasting robustness. We audit this distinction in frozen Chronos-Bolt and TimesFM forecasters on traffic occupancy and ATM withdrawals: 48 series per domain, 12 origins, and four nonzero delays. Forward-fill and a re-anchored forecast use the same measured history and predict the same targets; the latter forecasts across the gap before cropping its output. At the largest delays, forward-fill increases mean absolute scaled error by $41\%$–$88\%$ relative to re-anchoring, with positive paired effects in all four model–domain panels. Native coverage behaves differently across domains, so coverage alone does not diagnose the penalty. An exploratory traffic audit further reverses the apparent ordering of six- and 24-hour delays when forecast issue phase is averaged rather than fixed at midnight. We recommend separating unavailable information from its input encoding in robustness evaluations, without claiming a new forecaster or naturally observed delay distribution.