PATCH: Panel-Aware Hierarchical Conformal Cell Typing or Spatial Proteomics under Marker-Panel Shift

Cell phenotyping in spatial proteomics converts multiplexed tissue images into cell-type maps used to study tissue microenvironments, discover spatial biomarkers, and support patient stratification. It is hindered by cross-cohort shift in platforms and marker panels, and by the labor and subjectivity of manual gating. We present \textsc{Patch}, a deployment workflow with three components: a panel-aware hierarchical conformal layer, a label-free cohort-level shift diagnostic, and Vision Language Model (VLM) quality control. The conformal layer removes classes unsupported by the deployment panel and backs predictions off to the finest ontology ancestor identifiable from the observed markers; under exchangeability on a fixed panel, it inherits the standard split-conformal guarantee. Mean per-marker Kolmogorov--Smirnov distance, computed on unlabelled deployment cells, is used to decide whether zero-shot deployment is plausible or whether recalibration is needed. The VLM is applied only to ambiguous cells and may shrink the reported set, abstain, or flag artefacts, but cannot add unsupported labels. The underlying statistical head reaches $69.9\%$ flat accuracy and $57.5\%$ macro-F1 on the matched-input CRC30--40 benchmark. On $200{,}000$ held-out colorectal cells, \textsc{Patch} attains $0.98$ empirical ontology coverage at target $0.90$ with mean set size $|\mathcal{S}|=2.44$. Across held-out public cohorts, coverage decreases with mean per-marker KS distance; in a seven-cohort analysis, the correlation is $r=-0.85$ ($p=0.015$). Appending $5{,}000$ in-platform labelled cells restores target coverage on calibration-addressable cohorts, providing a practical deployment recipe under cross-cohort shift.