mirpy 4.6.0

Repertoire embedding now processes samples with bounded workers, retains their weighted RFF vectors and releases input frames. Shared clonotype-basis fitting selects rows without retaining the whole cohort. Batched scaler statistics, in-place standardization/PCA and bounded native embedding components reduce temporary memory while preserving SHM, row order and the established RFF representation.

Exports clonotype tables with sample/clonotype identifiers and unprojected sample RFF tables. Optional sample PCA can fit on those RFF rows, use compatible supplied cohort RFF rows, or apply a saved rotation. Existing signature corpus behavior and multi-locus MMD filenames remain unchanged. Requires vdjtools>=5.0.2.

Validation includes numerical/lifetime/worker regressions, the fast test suite, warning-free docs and isolated wheel checks. A generated 10,000-sample, seven-locus RFF table plus sample PCA stayed below 0.9 GiB peak RSS with one or four workers under a 16 GiB memory limit. Individual input depth and sample-table width still affect memory use.