Reference LoRA: synthetic shell mismatch can hide residual early cutoffs

Observed 2026-09-04 on ender-inference COR-3791, branch vansh/cor-3791-fine-tune-a-freight-turn-detector-for-identifiers-pauses-and, augmentation commit bc768b0. On a sealed balanced 42-row Reference holdout, the best weighted candidate’s seven false-finished cases included four bare partial-number prefixes followed by another number, while the 376-row paired synthetic augmentation contained zero bare-number-only shells. Two remaining failures only loosely matched a correction/restart stratum and one was outside all configured strata. Reusable method: before expanding synthetic data, map every residual error to an exact shell/target type; run an all-linear architecture comparison with the dataset and anchored objective held fixed, then—only if needed—add group-isolated paired bare-number partial/completed controls. Avoid teaching domain-specific adapters from unrelated semantic-continuation failures.