Supervise checkpointed Tower training

docs/CODEX_EXECUTION_LOG.md

@@ -11782,6 +11782,18 @@ Deployment evidence:
   supplied weights, runs calibration first, and rejoins the same automatic
   sampling/training path after rejection. Protected evidence remains closed
   until calibration passes.
+- The all-in-one container was externally recreated after v37 epoch 1. Both
+  456 MB checkpoints remained intact and training resumed from `last.pt` on
+  the RTX 4080 instead of restarting the experiment.
+- Added and activated a host-side YOLO supervisor. It requires the exact run
+  marker, a valid checkpoint larger than 1 MB, a running target container and
+  absence of `results.png` before issuing a bounded resume. It exits on a
+  completed training artifact or a missing/incomplete checkpoint.
+- The first live probe exposed that `docker top -eo args` is rejected by the
+  daemon and could misclassify an active Python-launched YOLO process. Two
+  transient duplicate resume processes were detected and terminated before
+  another epoch completed. Detection now uses `docker top -eo pid,args`; a
+  live one-shot check returned `monitoring` with exactly one GPU process.
 - Confirmed v37 epoch 1 completed on CUDA with validation precision `0.601`,
   recall `0.455`, mAP50 `0.474` and mAP50-95 `0.205`; the run remains inactive
   and these internal-validation metrics are not release evidence.
@@ -11792,6 +11804,8 @@ Verified in this pass:
   (`12 passed`).
 - `py -3 -m pytest -q backend/tests/test_belgium_training_loop.py backend/tests/test_failure_driven_yolo_sampling.py backend/tests/test_belgium_training_iteration_assessment.py`
   (`16 passed` after adding the completed-checkpoint entry contract).
+- `py -3 -m pytest -q backend/tests/test_yolo_training_supervisor.py backend/tests/test_belgium_training_loop.py`
+  (`11 passed`), plus a live supervisor one-shot and single-process GPU audit.
 
 Open:
 

docs/TODO.md

@@ -974,6 +974,7 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Allow the objective CUDA loop to consume an automatically clean `needs_human_review` corpus while keeping final human sign-off as a separate, mandatory promotion gate.
 - [x] Persist checksummed train-only failure-driven sampling after every rejected loop iteration and resume the next checkpoint from that exact dataset YAML.
 - [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] 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.