Support multi-sample promotion evidence

docs/CODEX_EXECUTION_LOG.md

@@ -5965,6 +5965,16 @@ Open:
   - the previous Geel 1024px upload `413` no longer occurs
   - the previous Geel low-threshold persisted YOLO/QA `504` no longer occurs
   - Geel `threshold=0.05` completed with 2,612 detections, 5,091 raw candidates, 2,479 suppressed duplicates, precision `0.06316998468606431`, recall `0.165` and F1 `0.09136212624584718`
+- Full AOI1024 positive persisted QA matrix:
+  - Geel best F1: `0.09136212624584718` at threshold `0.05`
+  - Turnhout best F1: `0.058721074894252295` at threshold `0.05`
+  - Retie best F1: `0.15621436716077539` at threshold `0.15`
+  - Westerlo best F1: `0.28703703703703703` at threshold `0.15`
+- Full AOI1024 hard-negative/background matrix:
+  - pure empty Postel/Lommel/Arendonk samples stayed at or near zero detections
+  - mixed background candidates still produced false-positive pressure: max detections were `59` at threshold `0.25`, `107` at `0.15` and `226` at `0.05`
+- Fixed `scripts/build_detection_model_promotion_report.py` after discovering it accepted `multi_sample_quality_summary.json` but counted positive samples as `0`. The report now supports both `samples[].runs[]` portfolios and `items[]` multi-sample summaries.
+- AOI1024 promotion report path: `artifacts/detection-model-promotion/aoi1024visible025e50-full/detection_model_promotion_report.md`; recommendation remains `none`.
 
 ## Known limitations
 
@@ -5972,7 +5982,8 @@ Open:
 - Several dense 1024 GRB reference exports reached the current 1000-feature source cap. Treat those samples as useful but potentially reference-capped until the provider query path supports paging or smaller dense AOIs are chosen.
 - The trained model is intentionally inactive. It needs persisted detection QA/QC matrix evidence and background/hard-negative evidence before default promotion.
 - The first completed persisted Geel QA run confirms the candidate is not promotion-ready: recall improves at low threshold, but false-positive pressure is too high.
+- The complete positive/background evidence confirms the candidate is not promotion-ready. The bottleneck is label/source quality and sample design, not runtime configuration.
 
 ## Next recommended pass
 
-- Run the remaining AOI1024 persisted QA samples and background/hard-negative checks, then decide whether label/source paging or additional AOI quality work comes before another training run.
+- Fix dense GRB reference completeness first: add provider-side paging or split dense AOIs so reference exports do not cap at 1000 features, then regenerate AOI1024 labels before another training run.