Rotate capped training samples across iterations

docs/CODEX_EXECUTION_LOG.md

@@ -11933,6 +11933,17 @@ Verified:
   AI-assisted inspection found labels aligned with visible roof footprints and
   retained the deliberately sparse hard-negative contexts.
 - NVIDIA preflight: RTX 4080 SUPER idle and available before the v43 handoff.
+- Iteration 2 selected its epoch-5 checkpoint, reached aggregate calibration F1
+  `0.584` and zero pure-empty detections, but remained closed because Flanders
+  F1/precision/recall and Wallonia precision failed. Iteration 3 started from
+  the exact rejected checkpoint without opening test/background.
+- Confirmed Ultralytics preserves duplicate paths from the 4,077-entry
+  failure-driven list. Added deterministic per-round rotation of the repeats
+  retained by the 65% regional cap, preventing persistent identical failures
+  from producing an identical capped list forever while preserving every
+  unique train tile and all protected-split exclusions.
+- `py -3 -m pytest -q backend/tests/test_failure_driven_yolo_sampling.py backend/tests/test_belgium_training_loop.py`
+  (`20 passed`).
 
 Open:
 

docs/TODO.md

@@ -1019,5 +1019,6 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Pass structural corpus and 56,474-label tile-quality audits with zero invalid or missing labels and zero spatial leakage failures.
 - [x] Render and inspect a 56-tile contact sheet covering all 14 new AOIs before retraining.
 - [ ] Run the v43 calibration-first, failure-driven CUDA loop against the frozen regional release gates.
+- [x] Rotate region-capped repeat windows deterministically per loop round so persistent failures cannot reuse an identical training list indefinitely.
 - [ ] Open protected test and pure-background evidence only after every calibration gate passes.
 - [ ] Queue final representative human sign-off only after all automated gates pass, then promote and redeploy the exact checksummed model.