Post-Extraction Compression of Pathology Foundation-Model Embeddings: Storage, Fidelity, and Downstream Utility

Whole-slide learning and retrieval often reuse large stores of patch embeddings extracted by a frozen pathology foundation model. We evaluate compression of 1,294,852 CONCH v1.5 embeddings from 1,701 TCGA glioma slides (877 patients) at $10\times$ magnification. We measure stored bytes, CPU decode time, reconstruction and retrieval distortion, attention stability, and three-class multiple-instance learning (MIL) under patient-disjoint evaluation. Per-channel INT4 achieves $8.95\times$ measured compression and a five-seed mean macro-AUROC of $0.8400$, compared with $0.8364$ for FP32. The patient-clustered 95% interval for the difference between five-seed ensembles is $[-0.0005,0.0164]$, above the specified $-0.01$ engineering margin. Adding one packed residual-sign bit per coordinate lowers mean attention total variation from $0.0305$ to $0.0166$ and increases top-5% attention-set Jaccard overlap from $0.921$ to $0.955$, while reducing compression to $6.92\times$. Normalization followed by Haar rotation reduces the largest recorded coordinate by $24.5\times$, but does not reduce maximum empirical excess kurtosis; the measured per-channel rotated decoder takes $3.57$ seconds per bag, compared with $22.28$ ms for INT4. We derive a bag-size-independent bound on gated-attention perturbation. These results support low-bit storage for this encoder, cohort, and MIL model while showing that attention fidelity and decode cost require separate evaluation.