Traced the gen4ve2e failure tissue mapping length (13) does not match the number of region classes (40) through SimStateBuilder and domain_from_histo.
Inspected two downloaded v2.0 domain representations. Their tissue metadata contains 40 declared classes: 28 base taxonomy IDs 0..27 and 12 grade-specific IDs 128..139.
Inspected the failing EE00009 segmentation. It contains 30 active values, including sparse IDs 129, 131, 132, and 133; the error reports all 40 declared metadata classes, not only active values.
Verified with git blame that Cedrik added the 13-entry _tissue_mapping() in commit 4f84c0c3 on 2026-08-14.
Lessons Learned: Pitfalls
Increasing downsampling_step cannot repair taxonomy metadata, so the runner’s generic retry repeats this deterministic failure.
Extending _tissue_mapping() from 13 to 40 entries is insufficient: domain_from_histo indexes the tensor by stored pixel ID, while v2 uses sparse IDs up to 139. The current region-info reader ignores the accompanying region_ids metadata and assumes dense IDs aligned with region_classes.
Lessons Learned: Improvements
Extend ZarrRegionInfo to read region_ids, validate a lookup table against those IDs, and construct composite names using mapping[region_id] rather than zipping class names with the mapping tensor.
Define and review a semantic mapping from all fine28 base and grade-specific classes into diffsim’s six tissue-characteristic groups. Do not assign new anatomy/artifact classes arbitrarily.
Treat taxonomy incompatibility as non-retryable so a resolution fallback is not attempted.
