Audit Belgian corpus splits and record CUDA evidence

docs/CODEX_EXECUTION_LOG.md

@@ -11467,3 +11467,34 @@ Next gate:
 - Verification: frontend build and 78 focused contracts passed. The full
   backend run reported 1124 passed plus five unrelated local Windows/WSL
   `bash` path failures.
+## 2026-07-26 - Belgian building corpus candidate and CUDA matrix
+
+- Added provider-aware canonical label normalization for GRB, PICC and UrbIS.
+  Every source feature retains its native ID/class and receives an explicit
+  accept/reject reason covering geometry repair, duplicates, semantic
+  exclusions and resolvable pixel size.
+- Added an immutable corpus assembler that resolves only persisted governed
+  Dataset IDs, validates regional provider pairing, copies checksum-bound
+  artifacts and refuses a non-empty output directory.
+- Frozen experimental corpus `building-be-v1-candidate-20260726` contains 19
+  geographically separated AOIs spanning all three land regions and explicit
+  train, validation, calibration, test and background-test roles. Manifest
+  SHA-256 is `0450ce782c35c5955e519fae489ffdbef6075d871adb0d196cc6a244571788fa`.
+- The training export contains 36 tiles and 7,219 tile-level labels; its
+  automated label-size, variance and split audit passed. A 12-tile regional
+  pilot contact sheet was rendered for human review.
+- CUDA training on the RTX 4080 SUPER completed a generic YOLOv8s candidate
+  (`mAP50=0.0883`, `mAP50-95=0.0248`) and an incumbent fine-tune
+  (`mAP50=0.187`, `mAP50-95=0.0617`) on the held-out regional validation
+  samples. Both are immutable candidates and neither is promoted.
+- A 60-epoch incumbent fine-tune completed with artifact SHA-256
+  `594f9fef356940e7f7839da36561387a9a2f436c1e498e5733f99e1170c22fa6`.
+  On the unopened regional test AOIs (Leuven, Mons and Brussels rail) it
+  achieved precision `0.315`, recall `0.229`, mAP50 `0.120` and mAP50-95
+  `0.0340`. The incumbent scored `0.227`, `0.233`, `0.0930` and `0.0262`
+  respectively. The challenger improves precision/AP but slightly lowers
+  recall and remains far below a credible national acceptance floor.
+- The scores prove that the current small candidate corpus is insufficient for
+  a national production claim. Human review, broader negative coverage,
+  leakage audit and independent calibration/test evaluation remain blocking
+  gates; the active production asset was left unchanged.