Dimensional accuracy in material-extrusion polymer additive manufacturing depends on feature geometry and process settings. This study evaluates a replicated, feature-resolved framework for polylactic acid (PLA) benchmark parts. A Taguchi L9 design varied layer height, print speed, and nozzle temperature across 45 independent builds, yielding 1575 measurements of 35 components across 16 features. Analyses combined build-level uncertainty, additive models, bootstrap intervals, exhaustive residual-permutation sensitivity, and clustered tolerance models. Inclined-surface roughness was summarized from five specimen means per condition. Mean absolute error (MAE) ranged from 0.0773 to 0.1263 mm. Layer height accounted for 91.8% of the decomposed condition-mean MAE variation (95% bootstrap interval: 90.0–93.2%). The additive-model effect had a Benjamini–Hochberg-adjusted p < 0.001; the nine-condition sensitivity analysis gave p = 0.093. Compliance odds ratios at ±0.10 mm were 0.51 and 0.32 for 0.2 and 0.3 mm layers relative to 0.1 mm. Nominally dominant factors persisted after dimension normalization. Dimensional utility, including slicer-estimated time, selected R2 (0.1 mm, 50 mm/s, 210 °C), while roughness-inclusive preferences made R1 competitive. Feature-specific tolerances and build-level uncertainty support process selection within the investigated PLA window.