Handwritten mathematical expression recognition (HMER) has traditionally converted handwritten images or ink into structured mathematical notation. Vision-language models (VLMs) now extend the task to correction, explanation, and grading. Yet their fluency can cause language priors to overwrite symbols, relations, or mistakes on the page. We systematically map 125 eligible studies from the January 2023-July 2026 frontier, reconstruct the benchmark and structural lineage, and verify claim-bearing VLM evidence through full-text examination. The evidence indicates that general-purpose VLMs broaden the task boundary without eliminating specialized visual and structural modeling. Handwritten-mathematics assessment benchmarks reveal a recurring failure chain in which perceptual errors become confident but unsupported explanations or repairs. Grading is most credible when specialized perception, rubric grounding, uncertainty, and human verification are treated as parts of the method. Evaluation must therefore extend from a single transcription to an evidence-preserving chain of recognition, transformation, and decision. We analyze how modality, representation, structural interfaces, supervision, and correction pathways determine whether downstream decisions remain visually grounded, and derive a three-part grounding contract. The resulting research blueprint combines encoder-decoder perception, tree/graph interfaces, and evidence-constrained VLM reasoning. It positions HMER as a testbed for trustworthy document intelligence and educational assessment.