Measurement-aware learning for reliable grain-boundary analysis in quantitative metallography

Quantitative metallography requires reliable boundary-derived measurement. Prior computational pipelines established feasibility, but reliability gaps remain when weak and interrupted interfaces are analyzed under crop-based workflows. Here, we study how context availability, annotation-aligned supervision, and measurement-oriented validation affect boundary-derived grain-size analysis. First, using a context-ablation experiment, we show that removing global context sharply degrades boundary identifiability, indicating that context is a first-order condition for reliable interface detection. As a methodological implementation of this measurement-aware framing, we introduce MLOGRAPHY++ , a context-preserving partial-label approach that aligns supervision with annotation ambiguity while reducing dependence on heavy post-processing. We then evaluate Heyn-Compare as an endpoint-oriented complement to pixel-overlap metrics for open-boundary predictions. Finally, we include diffusion-based super-resolution only as an exploratory representation stress test under fixed prompting, matched-resolution controls, matched detector training, and expert audit. Under these controls, discrimination metrics remain similar, while aggregate Heyn-based grain-size error decreases from 9.98 px to 5.38 px on the held-out split. These results support a bounded benchmark-level interpretation: reliable quantitative metallography benefits from coordinated attention to context, annotation reality, and measurement endpoints, while generated super-resolution detail should not be treated as physical high-resolution ground truth.