We revisit and substantially extend an early framework for Value-at-Risk (VaR) and Expected Shortfall (ES) of intrinsically quadratic portfolios of securities under multivariate elliptical and mixture-of-elliptical risk factors. The quadraticity considered here is conceptually different from the usual delta-gamma approximation of derivative portfolios: it arises directly from the second-order expansion of the exponential map linking log returns to security prices. This distinction yields a structured quadratic loss whose coefficients are determined by portfolio holdings rather than option Greeks. The revised analysis makes four substantive improvements. First, it separates the exact security loss, its linear approximation, and its quadratic approximation, and derives an explicit third-order remainder bound. Second, for long-only portfolios it identifies a canonical noncentral quadratic form and shows that the quadratic loss is bounded above by one half of initial portfolio value, thereby clarifying the local nature of the approximation and the origin of the classical restriction that the quadratic VaR remain below that bound. Third, it gives a clean joint characterization of VaR and ES through the distribution and lower partial first moment of a positive quadratic form. This yields transparent extensions to finite mixtures and precise existence conditions. Fourth, it develops modern computational routes for Gaussian and Student-t factors, combining generalized noncentral chi-square methods, scale-mixture representations, and Monte Carlo validation. A reproducible simulation study compares exact revaluation, linear approximation, and quadratic approximation under Gaussian, Student-t, and mixture risks. We then apply the framework to an equally weighted portfolio of eleven French equities observed weekly from September 2019 to September 2026. On the historical sample, the quadratic approximation is nearly indistinguishable from exact revaluation at the 95%, 97.5%, and 99% levels, while the linear approximation materially overstates tail risk. A fitted multivariate Studentt model yields an estimated degree of freedom close to 5.1, confirming substantial tail thickness. We further conduct an out-of-sample-style horizon validation using daily and directly observed weekly S&P 500, NASDAQ Composite, and CAC 40 returns from September 1 2016 to September 2026, together with an equally weighted three-index portfolio. Across all markets and confidence levels, the quadratic approximation remains extremely close to exact exponential revaluation; its relative error increases from daily to weekly horizons but stays below 0.4% even for weekly 99% ES, whereas the linear approximation can overstate weekly 99% ES by about 7%. We additionally compare directly observed weekly VaR and ES with the conventional √ 5 scaling of daily measures. The scaling rule is not uniformly accurate: it generally overstates weekly tail risk for the S&P 500, NASDAQ, and the equally weighted portfolio, while it can understate some CAC 40 weekly risks. The analysis therefore separates Taylor-approximation error from temporal-scaling error. The resulting framework preserves the original contribution while providing a more rigorous, transparent, and computationally modern treatment of quadratic portfolio tail risk.