mirpy-lib 4.5.1

Compute budgets now respect CPU affinity, scheduler allocation and container quotas through vdjtools 5.0.1. Embedding, density, clustering and ML runtime setup avoid nested numerical oversubscription. Shared cohort thread workers keep inner kernels serial.

Weighted density chooses blocks from actual neighbour counts, keeping sparse queries efficient and dense memory bounded while retaining exact per-row reduction order. Clustering preserves relative negative sklearn worker counts within the allocation.

Requires vdjtools>=5.0.1. Reference models and the embedding coordinate system are unchanged.

Copilot findings are fixed. All five CI jobs passed on the reviewed source, including optional Torch, ANN and BioPython coverage; 63 focused runtime/density/clustering tests and warning-free documentation also pass.