Contextual stochastic optimization (CSO) uses covariates to guide decisions under uncertainty, yet its finite-sample behavior is still not well understood. We study two common frameworks for CSO: sequential learning and optimization (SLO) and integrated learning and optimization (ILO). We first show that the estimation-prescription mismatch persists even under stringent pairwise accuracy comparison criteria: a more accurate estimate can induce a worse decision. This motivates the prescriptive curve, which tracks the exact expected out-of-sample cost as a function of training data size under SLO or ILO. We construct contextual linear and nonlinear optimization problems in which these curves strictly increase over a finite range of data sizes under both frameworks, even when the hypothesis class is well-specified. Given this counterintuitive finding, we show that one more data observation is not universally beneficial in the finitesample regime. A single finite-precision sample can be compatible with data-generating processes that imply opposite signs for the one-step change in the prescriptive curve when the training data size increases from N to N+1. We next examine jackknife aggregation as a conservative way to use one additional observation: relative to the traditional procedure trained on one fewer observation, it can strictly improve expected prescription performance under some data-generating processes and strictly worsen it under others. Finally, we extend the analysis beyond SLO and ILO: a universally consistent but finite-sample sub-optimal data-driven policy need not admit another universally consistent policy that, with one additional observation, is never worse across all distributions in an ambiguity set and all finite data sizes, and is strictly better for at least one distribution-sample-size pair.