We study coherent tail-risk measurement under distributional ambiguity and ask how a robust VaR-Expected Shortfall forecast should be statistically validated. The structural part of the paper merges model ambiguity and ES tail aversion into an effective dual envelope, relates strong time consistency to rectangularity, and constructs the minimal rectangular extension of a nonrectangular ambiguity set. The econometric contribution departs from pointwise backtesting. We define the ambiguity risk region as the image of a distributional ambiguity set under the jointly identifiable VaR-ES functional and formulate calibration as a set-membership problem. For one-dimensional W 1 balls of radius c/ √ m, the induced region is nonregular: its VaR width is O(m-1/4), whereas its ES width is O(m-1/2). We derive the corresponding deterministic mixed-rate limit set and show that this geometry generates a three-regime out-of-sample test. If the local VaR displacement satisfies |h| < a c , with a c = (2c/f 0 (q 0)) 1/2 , the VaR moment is profiled out and the limit is the squared distance of a Gaussian variable to an ES interval. At |h| = a c , the statistic converges to a Gaussian cone-projection law and, under the refined boundary null, to 1 2 χ 2 1 + 1 2 χ 2 2. If |h| > a c , the statistic diverges at rate √ m, with an explicit leading coefficient. Distributional coverage of the ambiguity ball is shown to imply coverage of its VaR-ES image. An empirical application to daily CAC 40, S&P 500, and NASDAQ losses over 2016-2026 places all three markets in the interior regime at the 97.5% level. In each case the 2023-2026 empirical distribution lies outside the calibrated training W 1 ball, while its VaR-ES pair remains inside the induced risk region. This distinction illustrates the practical content of ambiguity calibration: distributional incompatibility need not imply tail-risk incompatibility.