This paper presents a complete predictive framework for the "lockstep deficit" – the systematic failure of the standard Hardy–Littlewood independence heuristic to correctly estimate the probability that the immediate successor of a prime gap starter is itself a starter. We define an h-starter as a prime p such that p+h is also prime, and L_h(X) as the fraction of h-starters p ≤ X whose very next prime is also an h-starter. The naive heuristic predicts that L_h(X) should equal S(h)/ln X, where S(h) is the exact singular series constant for shift h. Using direct sieves to X = 2×10^8, covering over 11 million primes and multiple shift classes, we show that this prediction fails, but that the failure is completely and quantitatively explained by a conditional singular series with zero free parameters. Three main results are established. First, we derive an exact normalization identity: the commonly used endpoint ratio L_h(X)/(S(h)/ln X) overstates the true conditional density by the factor ln X · E[1/ln q], where q ranges over the immediate successors of starters. This algebraic identity, verified numerically to five decimal places, resolves the apparent factor-of-five discrepancy reported in earlier analyses and shows that the true object of study is the per-event ratio R^ev_h. Second, the same conditional-density model, applied to nine distinct shifts h in {2, 6, 10, 30, 90, 144, 210, 288, 720}, predicts correction factors spanning two orders of magnitude – from an exact zero at h=2 through 0.20 at h=10 to 0.965 at h=210 – and matches direct measurement in every case, with all residuals lying between -0.3% and +2.9%. This cross-shift validation is the paper's strongest test: shifts with identical S(h) constants (e.g., 6, 144, 288) nevertheless exhibit widely different lockstep corrections, and the model captures these differences entirely from the interaction between h and the empirical successor-gap distribution. Third, we demonstrate that the deficit is not a constant. Because the correction factor averages over the widening distribution of prime gaps, it drifts with X. The model reproduces this drift quantitatively: across X from 10^7 to 2×10^8, the derived correction rises from 0.9086 to 0.9150, while the observed value rises from 0.9132 to 0.9231, with a stable residual of approximately +0.7%. We additionally prove two exact-zero theorems: (i) lockstep is impossible for h=2 (confirmed by zero successes among over 800,000 twin leaders), and (ii) a gap of exactly 2h between consecutive h-starters is forbidden for every even h (verified for five representative shifts). The remaining ~1% residual is localized to large successor gaps, consistent with the sole unmodeled effect – conditioning on the fact that no prime exists strictly between p and p+g (interval avoidance). This is stated as the central open problem, with a Buchstab-type refinement proposed as the natural path to closure. All computations use deterministic sieves of Eratosthenes, independently verified across three passes and hardware configurations. Euler products for singular series are evaluated with the inadmissibility convention (zero when a constellation covers all residues modulo any prime). Full reproduction scripts and raw data tables are archived with the release.