portable runner: sys.path bootstrap + concrete precision resolution

scripts/run_stage_portable.py

@@ -37,10 +37,17 @@ import math
 import os
 from pathlib import Path
 import shutil
+import sys
 import tempfile
 import time
 import uuid
 
+# The pear15 package sits next to scripts/ (repo root locally, CODE/ on Kaggle).
+_HERE = Path(__file__).resolve().parent
+_REPO_ROOT = str(_HERE.parent)
+if _REPO_ROOT not in sys.path:
+    sys.path.insert(0, _REPO_ROOT)
+
 TARGET = 50_000_000_000
 MAX_SESSION_SECONDS = 43_200
 FORMAT = "pear15-portable-stage-runner.v1"
@@ -153,7 +160,12 @@ def run(mode, head_root, stage_root, data, *, seconds=7200, device="cuda",
                  if gradient_accumulation_steps is not None else cfg["gradient_accumulation_steps"])
         grad_ckpt = (gradient_checkpointing
                      if gradient_checkpointing is not None else True)
-        resolved_precision = precision or "auto"
+        # Resolve now (idempotent with the trainer's own resolution) so lineage,
+        # stage-head and the saved config all carry a CONCRETE precision. An
+        # 'auto' recorded verbatim would make the next session think the stage
+        # is unresumable and silently migrate again.
+        resolved_precision = core._resolve_precision(
+            torch, device, precision or "auto")
         chunk = loss_chunk_tokens if loss_chunk_tokens is not None else min(
             cfg["loss_chunk_tokens"], micro * seq)
         cfg.update(