Stabilize near-gate regional sampling

docs/CODEX_EXECUTION_LOG.md

@@ -11944,6 +11944,12 @@ Verified:
   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`).
+- Iteration 3 raised aggregate calibration F1 to `0.605`, passed Brussels and
+  every Wallonia gate, and isolated the remaining blocker to Flanders. When
+  iteration 4 removed Wallonia stabilization, Wallonia precision regressed
+  narrowly from `0.525` to `0.496`. Added a `0.03` precision/recall sampling
+  guard-band so a just-passing region retains stabilizing evidence while the
+  hard-failing region remains the primary target (`21 passed`).
 
 Open:
 

docs/TODO.md

@@ -1020,5 +1020,6 @@ This file now starts with the current implementation status. Older preparation/b
 - [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.
+- [x] Keep near-gate regional precision/recall stabilization inside a 0.03 sampling guard-band to prevent cross-region seesaw regressions.
 - [ ] 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.