Battery state-of-health (SOH) benchmarks often mix cycles from the same cell across training and test sets. We examined whether predictor structure changes the gap between within-cell interpolation and cross-cell transfer. Seven regressors and five predictor sets were evaluated in four NASA cells (636 cycles) using ten random 80:20 splits and leave-one-cell-out (LOCO) testing repeated over ten model seeds. Across-model median random split RMSE fell from 5.00 SOH points with age alone to 1.52 with diagnostics plus age, whereas median LOCO RMSE fell only from 5.83 to 4.61; the absolute protocol gap therefore increased from 0.83 to 3.09 points, and the median cell-paired LOCO/random ratio increased from 1.24 to 4.53. With all diagnostics, model-level ratios ranged from 1.65 to 5.16, and random-versus-LOCO ranks were uncorrelated (descriptive Spearman ρ = 0.00). Holding the number of training cycles constant, increasing the number of source cells from one to three reduced median LOCO RMSE from 5.25 to 4.73 points, but 32.1% of cell–model pairs did not improve. Standardised sliced Wasserstein shift was associated with the validation ratio (descriptive ρ = 0.73) but not with absolute LOCO RMSE (ρ = −0.01). Rated capacity normalisation preserved the predictor-dependent gradient. An eight-cell Oxford analysis was limited to an age-only boundary check and was not a feature-matched replication. These four-cell laboratory results show that richer discharge diagnostics can improve represented cell interpolation much more than unseen cell transfer; claims should therefore report absolute and relative errors with complete cell-held-out validation.