Replaced every NanoVirDx *Batch struct-of-lists with scalar *Sample dataclasses. Predictors now consistently accept and return lists while preserving input order and top-level cardinality.
Preserved the intentional one-to-many Histo preprocessing relationship as PreprocessingOutputSample.components: list[WSISample].
Moved fine28 mapping, nine-tissue parameters, sparse Histo v2 region loading, and DiffSim 0.1.100 compatibility into inference/simulation/_compat.py. Reduced diffsim.py to simulation sequencing and output translation.
Finished the user’s vxData adapter extraction. Both get_*_by_uid functions accept an optional client, instantiate their own fallback only for non-empty inputs, preserve requested ordering and duplicates, handle local/S3 paths, and include nuclei maps.
Added a NanoVirDx conda recipe and manual release workflow. The workflow reads the hardcoded pyproject.toml version and uploads only to the internal Artifactory virdx channel.
Removed stale README examples for APIs that are not implemented and documented the list-of-samples contract.
Committed and pushed mono commits 791be45a and 67a029d2 to origin/fk/nanovirdx.
Added ISUPGradePredictor and PICAIPredictor placeholder interfaces whose setup() and predict() methods explicitly raise NotImplementedError; pushed mono commit 1d6c28c5.
Added Gleason, tissue, and expanded-cell component predictors as projections of the Histo v2 domain pipeline; removed the unsupported WSIEmbeddingSample placeholder because Histo has no persisted FeatureGrid contract; pushed mono commit 1b37a2b1.
Added EmbeddingGenerator writing a slide-level H-optimus token mosaic with valid-core ownership over Histo’s super-window lattice; reinstated WSIEmbeddingSample(path) as a NanoVirDx-owned Zarr artifact contract; pushed mono commit 26f2a957. Correction: Histo has a stable in-memory FeatureGrid API; only the persisted format was missing.
Rebuilt Gleason, tissue, and cell predictors on Histo’s official head pathways (load_gleason_head/run_gleason_segmentation, load_tissue_head/run_tissue_segmentation, load_cell_head/run_cell_detection) via a shared WsiHeadSweep over Histo’s super-window lattice; renamed the settings bridge to _compat_dr_v2_config.yaml; typed settings as PipelineSettings; pushed mono commit bb61748a.
Added explicit nanovirdx.inference.__init__ exports (from nanovirdx.inference import DomainRepresentationPredictor, ISUPGradePredictor, ...) plus empty subpackage __init__.py files; heavy runtimes stay lazy in setup(); pushed mono commit 0330a4d1.
Lessons Learned: Pitfalls
Running independent workers in one shared worktree requires strict file ownership. The broad sample migration worker therefore produced a read-only design while the other workers owned only new, non-overlapping files.
Domain-representation downloads need one coherent directory per resource because DiffSim discovers conventionally named maps from a directory. Downloading multiple representations into one flat destination risks collisions.
Empty predictor inputs should return before model setup; otherwise predict([]) can needlessly load large Histo or DiffSim state.
Lessons Learned: Improvements
Add a small NanoVirDx project/test target so the scalar sample and optional-client contracts run in CI rather than only through focused validation scripts.
Remove _compat.py after DiffSim publishes Cedrik’s region_ids-aware name mapping support.
Remove the vendored Histo settings after Histo packages a supported default-settings resource.
