Continue rejected v37 through frozen loop

docs/CODEX_EXECUTION_LOG.md

@@ -11804,6 +11804,19 @@ Deployment evidence:
   malformed JSON and filesystem errors produce an explicit fail-closed
   supervisor state before any subprocess is started. The updated supervisor
   was activated live with one remaining v37 GPU process.
+- V37 early-stopped normally at epoch 23; epoch 5 remained the immutable best
+  checkpoint. The supervisor observed `results.png` and executed its one-shot
+  completion handoff.
+- Fixed a completed-checkpoint entry defect exposed by that live handoff: the
+  iteration directory is now created before copying/hash-binding an existing
+  model. The repaired calibration ran without opening test/background.
+- V37 improved calibration aggregate F1 to `0.568` and Flanders F1 to `0.307`,
+  but failed Flanders F1/precision/recall and Wallonia precision. It was
+  rejected and produced checksummed failure-driven sampling for iteration 2.
+- Made training patience an explicit orchestrator input. A pre-epoch iteration
+  2 process exposing the old hardcoded value `35` was terminated before any
+  result row existed; iteration 2 restarted from the identical checkpoint and
+  sampling with the frozen `patience=18` contract on CUDA.
 - Added those aerial augmentation parameters to the orchestrator CLI and
   training command, preventing later failure-driven checkpoints from silently
   reverting to generic orientation assumptions.

docs/TODO.md

@@ -976,6 +976,8 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Add a guarded calibration-first entry point for completed checkpoints so v37 and future externally interrupted runs can rejoin the automated loop without redundant retraining.
 - [x] Add and activate a host-side, exact-run-marker supervisor that resumes the v37 CUDA checkpoint after container recreation without launching concurrent trainers.
 - [x] Bind v37 completion to a one-shot JSON command that enters calibration-first evaluation and up to 20 automatic failure-driven CUDA iterations with the frozen aerial augmentation contract.
+- [x] Complete v37 by early stopping at epoch 23, reject its best epoch-5 checkpoint calibration-first, and start closed-loop iteration 2 from checksummed failure-driven sampling.
+- [x] Make orchestrator patience explicit and retain the frozen value `18` in every subsequent CUDA iteration.
 - [x] Evaluate the completed v36 YOLO11x checkpoint calibration-first on the rotated v30 holdouts; reject it before opening test/background because the regional calibration gate failed.
 - [ ] Finish and assess the leak-free v37 YOLO11x failure-driven CUDA iteration; open test/background evidence only if every calibration gate passes.