Add v44 Flemish corpus expansion

docs/CODEX_EXECUTION_LOG.md

@@ -11950,6 +11950,12 @@ Verified:
   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`).
+- Iterations 5 and 6 confirmed that the guard-band retains Wallonia above its
+  gates, but Flanders remains structurally below precision, recall and F1
+  gates. Stopped iteration 7 before its first result row and freed the GPU
+  rather than repeating the same evidence. Added 12 independent v44 Flemish
+  acquisition targets across coastal, industrial, ribbon, dunes, port and
+  farmland-hard-negative contexts for the next immutable corpus wave.
 
 Open:
 

docs/TODO.md

@@ -1019,6 +1019,7 @@ 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.
+- [ ] Acquire and integrate the v44 Flemish-only corpus wave after six loop assessments isolated coastal, industrial and ribbon-domain under-generalisation.
 - [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.

scripts/provision_belgium_building_training_portfolio.py

@@ -102,6 +102,22 @@ AOIS = (
     Aoi("veurne-ribbon-train-v42", "flanders", "ribbon-development", "train", 2.666, 51.073),
     Aoi("zwin-dunes-train-v42-bg", "flanders", "dunes-negative", "train", 3.360, 51.350, "background_candidate"),
     Aoi("nieuwpoort-port-train-v42-bg", "flanders", "port-hard-negative", "train", 2.735, 51.146, "background_candidate"),
+    # v44 corpus correction after six closed-loop assessments isolated the
+    # remaining blocker to Genk/Lokeren/Ostend-like Flemish roof domains.
+    # These centers are independent from the protected calibration footprints
+    # and expand real evidence rather than repeating the same capped tiles.
+    Aoi("middelkerke-coastal-train-v44", "flanders", "coastal-urban", "train", 2.820, 51.185),
+    Aoi("heist-coastal-train-v44", "flanders", "coastal-urban", "train", 3.245, 51.338),
+    Aoi("westende-coastal-train-v44", "flanders", "coastal-urban", "train", 2.768, 51.170),
+    Aoi("dessel-industry-train-v44", "flanders", "industrial", "train", 5.112, 51.238),
+    Aoi("olen-industry-train-v44", "flanders", "industrial", "train", 4.858, 51.143),
+    Aoi("evergem-industry-train-v44", "flanders", "industrial", "train", 3.710, 51.113),
+    Aoi("waasmunster-ribbon-train-v44", "flanders", "ribbon-development", "train", 4.085, 51.110),
+    Aoi("wingene-ribbon-train-v44", "flanders", "ribbon-development", "train", 3.274, 51.058),
+    Aoi("kasterlee-ribbon-train-v44", "flanders", "ribbon-development", "train", 4.966, 51.241),
+    Aoi("de-panne-dunes-train-v44-bg", "flanders", "dunes-negative", "train", 2.575, 51.112, "background_candidate"),
+    Aoi("zeebrugge-dock-train-v44-bg", "flanders", "port-hard-negative", "train", 3.190, 51.345, "background_candidate"),
+    Aoi("diksmuide-field-train-v44-bg", "flanders", "farmland-hard-negative", "train", 2.830, 51.020, "background_candidate"),
     Aoi("bruges-val", "flanders", "historic-urban", "val", 3.224, 51.209),
     Aoi("turnhout-val", "flanders", "suburban", "val", 4.944, 51.322),
     Aoi("hasselt-cal", "flanders", "suburban", "calibration", 5.340, 50.930),