1.0.1

This patch release brings tile-encoder extraction into line with the authors' own preprocessing, adds Mettle as a tile encoder, and makes shared-k benchmark caches safe when the roster changes. It also ships a pooling-sensitivity study and a much expanded shortcut-susceptibility (nIPD) section on the documentation site.

Changed

• Six tile encoders now use their authors' preprocessing recipes. DINOv2-B, H-optimus-0, H-optimus-1, Prov-GigaPath, GPFM and mSTAR used to take their transform from timm's hub config. Each now has an explicit resize/crop/normalisation recipe that matches slide2vec, and that recipe is recorded in the artifact contract. Embeddings extracted earlier for these six models no longer match the contract and must be re-extracted. The other 21 tile encoders are unchanged. (#199)
• The six models' published results were regenerated under the new recipes. That covers Camelyon, TCGA-4×4, Tolkach-ESCA and their nIPD cells. DINOv2-B now reproduces FMTF's slide2vec run, and mSTAR's Camelyon margin turns negative, which is a real robustness difference. (#200)
• Median-k caches are keyed by the resolved shared k. When adding a model moves the shared k, the affected RI/MaRI outputs are recomputed instead of being served from a stale cache. k-NN, RI-curve and CRoMa caches stay independent of k. Headline biological and confounder balanced accuracy are now reported at the same operating k. (#170)

Added

• Mettle tile encoder (slideflow-labs/Mettle, pinned revision, FP32, 3072-d cls_mean view). The loader checks the checkpoint hash and the processor contract, and an opt-in smoke test covers it. (#196) Mettle is now published in the tile panel, ranking fifth of 26 encoders; it leaves the shared k unchanged on every cohort. (#202)
• Pooling-sensitivity study. It compares canonical against alternative pooling for Mascaret, Phaet and the RudolfV-2 family across the four PathoROB benchmarks, with deterministic paired bootstrap and protection for the canonical artifacts. All five encoders keep their canonical pooling, and Mascaret stays the only frontier member. (#171, #172, #176, #177, #178)

Documentation

• Shortcut-susceptibility (nIPD) results are published on the site:
  • a static page (#187)
  • an interactive cohort-first explorer (#188) with two-model comparison (#189)
  • a linked CRoMa–nIPD scatter (#190)
• The section was renamed and reorganised. The page is now Shortcut susceptibility; the old URL redirects. Each cohort page carries its own nIPD tables. baseline_skill and the docs-side Δ(MaRI−RI) were removed, and the nipd.json schema is now v2. (#191)
• nIPD tables surface and rank on the change at V = 1. The site shows that endpoint next to nIPD and flags curves that end at chance (#192). The tables now rank on it, because nIPD's signed area can cancel and hide a late collapse (#193).
• Readers can request a tile or slide encoder from the docs through new GitHub issue forms. (#179, #180)
• Row shading. Data-overlap shading now appears on every cohort page; RI-in-development rows get their own yellow tint, and each table caption states its k and roster size. (#202)