arXiv:2610.19229v1 Announce Type: new Abstract: Paper 1 established the per-bundle statistical limits of Temporal CT (TCT) and benchmarked a learned prior on a single patient at a uniform incident count $N_0$. This paper lifts both restrictions. SNN2, a Gauss-Newton estimator, converges to the per-bundle Poisson maximum-likelihood solution and is the prior-free baseline. N$_0$FC, a bow-tie and tube-current-modulation co-design, compresses within-bundle photon-flux dynamic range. V8c is a neural-network residual corrector trained on a 19-patient corpus stratified by water-equivalent diameter (WED). T2bp, a per-patient empirical-Bayes post-processor, combines per-bundle Poisson likelihoods with a prior built from the patient's own bundles; it requires no training, so it cannot inherit a mismatched training prior. Results use Paper 1's bias-charged Dose Inflation Factor (DIF), extended to per-ray budgets. V8c beats SNN2 on in-distribution patients under baseline acquisition but fails out of distribution. Composed with SNN2, T2bp matches or beats T2bp$\circ$V8c on three of four N$_0$FC test cells (WED 207-342 mm), with the largest gains on the high-WED and out-of-distribution patients, and trails by 3% on the fourth. On 23 non-test corpus volumes, N$_0$FC with closed-loop tube-current modulation lowers the per-volume median middle-path Cramer-Rao floor, as a DIF, from 23-48 under baseline acquisition to 3.3-6.3, near the equal-attenuation value 13/3; T2bp$\circ$SNN2 reaches this floor without training, its mean squared error 0.74-1.26 times the floor (median 0.89) over the bulk of each volume. Acquisition co-design removes most of the attenuation-contrast penalty, and a training-free estimator attains the resulting limit. The remaining three- to six-fold gap to dose parity with single-source CT, and the high-attenuation tail, are the target of inter-bundle priors in later work.