We propose a coherent probabilistic model for consecutive prime gaps inside a fixed arithmetic progression modulo M. The model combines a Cramér-type intensity filtered by the residue class, Hardy—Littlewood two-point correlations via the singular series S(g), and an exponential suppression of intermediate primes. Under natural assumptions, the relative frequencies of small admissible gaps satisfy freq(g2)/freq(g1) ~ S(g2)/S(g1). We test the model on the first 10^6 primes (up to 1.5×10^7) in the digital root classes modulo 9 using exact rational arithmetic in SageMath. The predicted resonance for gap 90 (excess factor 4/3) is observed as +34.5% against +33.3% predicted, an agreement within 1.2 percentage points over 166,567 gaps. For gaps approaching the mean spacing the two-body approximation breaks down, with a sign inversion at gap 198, clearly marking the transition scale. All computations confirm the structural Lemma 1 (zero violations) and the asymptotically stable product freq(g_min) × mean_gap ≈ 2 C2 S(g_min) φ(M).