TODO.md

Primary

• Recreate LazyTransform from uns
• Numerical precision inconsistency
  • Mix of float32 and float64 (mainly in operator, sometimes in-mem) causes drift within and across modes
  • C++ uses 64-bit, R/Rcpp does also — always consistent
• Fix R/Python result parity
• Change output from numpy to pandas (annotate_*)
• Reorganize and consolidate code
  • Audit for dead functions
• Interactive cell selector
• UMAP points have no stroke
• Remove violin boxplots? Quantiles don't align.
• compute_archetype_feature_specificity(): key_added > key_prefix

Secondary

• Consolidate normalization code-paths
• Add network centrality to run_actionet?
• Lazy transform in-memory?
• Allow changes to lazy transform params
  • Force recompute of attributes
• lazy_logcounts in _validate_lazy_logcounts_params() does nothing?
• compute_transparency() use scale()
• Explore more accurate and faster log approximations for JSD and lazy transform
• Combine plot_umap* paths with raster=bool
• Python compute_feature_specificity() in-mem is still garbage (maybe not?)

Done

• ARMA_DONT_USE_WRAPPER multiple redefinition
• Parallel specificity bug (fixed???)
• Standardize key args
• Test impute features
• Implement plotFeatures
• Test backed SVD
• MatrixSource supports layers, but layers can't be backed.
  • Added validation logic
• Add pseudocount to normalize_anndata()
• compute_feature_specificity backed/in-mem parity
• Fix memory usage in compute_network_diffusion() (in-memory)
• Speed up plot/impute features
  • I/O bound. Probably as good as it's gonna get
• Add run_lpa/propagate_labels
• Decouple archetype specificity from run_actionet
  • Compute specificity on the fly for impute_from_archetypes()
• Decouple compute_network_diffusion from anndata
• Add 3D to plotly
• Optionally omit C_* and specificity matrices to reduce object size
• Document everything (OpenMP hard requirement, README overhaul, context files updated)