Recover active training loop checkpoints

docs/BELGIUM_BUILDING_TRAINING_LOOP.md

@@ -106,6 +106,13 @@ versioned JSON argv list. The handoff starts the orchestrator detached exactly
 once; shell strings are not accepted. Subsequent iterations retain the frozen
 180-degree aerial rotation, vertical/horizontal flip, scale and translation
 parameters rather than silently reverting to generic augmentation defaults.
+Before fitting a state-pending iteration, the orchestrator checks its canonical
+run directory for `weights/last.pt`. If present, it resumes that exact CUDA
+checkpoint instead of restarting from the prior candidate. A host-side parent
+supervisor monitors the exact loop-process marker and persisted loop state; on
+container recreation it restores the current orchestrator script and launches
+the versioned JSON argv command. Invalid state and bounded launch exhaustion
+fail closed.
 The orchestrator refuses to start unless every automated frozen-dataset gate
 passes and the corpus contains zero blank/low-variance positive tiles. The
 audit status may remain `needs_human_review` while training and objective

docs/CODEX_EXECUTION_LOG.md

@@ -11827,6 +11827,15 @@ Deployment evidence:
   cleanly from the same candidate and checksummed sampling with
   `warmup_epochs=1`, `warmup_bias_lr=0.01`, HSV `0.01/0.2/0.15`, patience 18
   and the existing aerial rotation/flip contract.
+- Added checkpoint-aware recovery for the state-pending current iteration.
+  When `runs/<iteration>/weights/last.pt` exists, the orchestrator uses YOLO's
+  exact `resume=<checkpoint>` CUDA path and records that provenance after the
+  iteration is assessed.
+- Added and activated a host-side parent-loop supervisor. It monitors the exact
+  orchestrator marker and fail-closed JSON loop state, restores the current
+  script into a recreated container and relaunches only the versioned argv
+  command. Its live state is `monitoring`, with zero relaunches and one active
+  iteration-2 GPU process.
 - 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

@@ -979,6 +979,7 @@ This file now starts with the current implementation status. Older preparation/b
 - [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] Preserve v37 warmup and aerial HSV augmentation in every loop iteration; archive and invalidate the pre-epoch-2 run that exposed generic Ultralytics warmup drift.
+- [x] Resume a partially written current loop iteration from its exact `last.pt` after container recreation and activate a host-side parent-loop supervisor.
 - [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.