Inspected the exact histo-v0.0.127 source for cell detection, tissue and Gleason heads, domain-representation model loading, whole-slide processing, and output persistence.
Confirmed that process_scans is the stable whole-slide orchestration boundary. It reuses the watershed pool and host reservations across super-windows and scans.
Recommended that NanoVirDx keep a path-backed WSI-to-five-Zarr predictor contract and delegate the whole input list to Histo instead of exposing FeatureGrid or duplicating super-window logic.
Found that the current NanoVirDx wrapper calls process_scan without reusable watershed or reservation objects, which starts cell post-processing resources per super-window.
Lessons Learned: Pitfalls
The installed macOS Pixi environment does not contain Histo because the dependency is Linux-only. The local Histo repository and exact release tag were the reliable source.
Standalone cells_zarr persistence is not the domain-representation v2 persistence path; v2 writes raster instance maps and a cell lookup table through WsiOutputs.
Lessons Learned: Improvements
Document process_scans as the preferred external inference adapter, including its output-directory layout and completed-scan skip behavior.
Document that process_scan callers must supply a shared CpuWatershed and host reservation when processing multiple super-windows, or accept substantial repeated startup/allocation overhead.
