Primitive-root determinant densities: extremal order, dimension-zero, Fourier decay, and a lattice-smoothing no-go

For a prime $p$, let $c(p)$ denote the limiting fraction of $n\times n$ matrices over $\mathbb{F}_p$ whose determinant is a primitive root modulo $p$. The quantity $c(p)$ is a natural multiplicative deformation of the totient ratio $φ(p-1)/(p-1)$ and inherits its distributional behaviour over the primes. Existence and continuity of the limiting law follow from the shifted-prime Erdős--Wintner--Hildebrand framework. We prove the following new results: unconditionally, $\inf_p c(p)=0$ and the sharp order is $\min_{p\le x}c(p)\asymp 1/\log\log x$; the reciprocal satisfies $\limsup_{p\to\infty, \, p\text{ prime}} 1/(c(p)\log\log p)=e^γ$, and no smaller constant suffices. We give a complete proof, combining an adaptation of Erdős's argument with the Jessen--Wintner pure-type dichotomy, that the limiting distribution is purely singular, and strengthen this to $\dim_H(μ_G)=0$, i.e. the limiting measure is carried by a Borel set of Hausdorff dimension zero. The distribution has full topological support $[0,\tfrac12]$ and admits a Bernoulli-product representation indexed by the odd primes. The moments are given by convergent Euler products, and the Mellin transform $\mathbb{E}[X^s]$ extends to an entire function of $s$, non-vanishing on $\operatorname{Re}(s)>0$. Near the right endpoint, $1-G(\tfrac12-\varepsilon)\simκ/\log(1/\varepsilon)$ with an explicit constant $κ$. The quantity $1/c(p)$ equals the dimension-uniform expected rejection-sampling overhead in the reduction from learning with errors (LWE) to PRIM-LWE in lattice-based cryptography. The explicit bounds yield concrete overhead estimates for all primes appearing in current NIST post-quantum standards and representative NTT-friendly moduli.