Make dataset audit mandatory for training loop

docs/CODEX_EXECUTION_LOG.md

@@ -11570,3 +11570,20 @@ Next gate:
   became Brussels `0.566`, Wallonia `0.427`, Flanders `0.222`. The generic-base
   iteration did not beat it. Both remained rejected, and iteration 4 started
   from the stronger candidate on the dated 75-AOI corpus.
+
+## 2026-07-27 - Positive-imagery gate and complete SPW campaign
+
+- Visual inspection caught 20 positive Walloon tiles with labels over white
+  no-data imagery. SPW 2024 is an official but partial campaign, so that
+  training run was stopped and its candidate is invalid.
+- Dataset QA now fails whenever any positive tile is blank/low-variance. The
+  checkpointed training orchestrator also refuses to start unless this audit
+  is `ok` with zero affected positive tiles.
+- Replaced the partial product with the official, territory-complete SPW summer
+  2023 campaign (27 May through 25 June, 25 cm). The frozen replacement corpus
+  `building-be-v5-dated-20260727-r1` retains 75 AOIs and 13,765 accepted labels.
+  Manifest SHA-256 is
+  `eade90d3ba22b426b72a300fcaa9a01f6eb1d18567c4a53413b72b8048f5f021`.
+- The v5 tile audit passed with zero blank positive tiles. The automated CUDA
+  loop started from the strongest prior candidate and will checkpoint every
+  train/calibrate/test/background assessment without promoting failed models.