preflight: P2 fast-fails after the first dead cell (each repeat bills a GPU session)

train/preflight.py

@@ -326,6 +326,14 @@ def p2_throughput(args, R):
                 cell["tail"] = r["out"][-1500:]
                 cell["err"] = r["err"][-800:]
             out[f"L{layers}_s{seq}"] = cell
+            if r["rc"] != 0:
+                # Fast-fail: a systematic defect (import, OOM, batch-shape) repeats identically in all
+                # four cells, and each repetition bills a GPU session. Stop at the first one, report it,
+                # fix, then re-run. Cheapest possible way to be wrong.
+                R["P2_aborted_after"] = f"L{layers}_s{seq}"
+                break
+        if R.get("P2_aborted_after"):
+            break
     R["P2"] = out
     got = [c for c in out.values() if isinstance(c.get("tok_per_s"), (int, float))]
     # Pass means all four cells ran AND the frozen cell (22L/seq2048) is reported, because that is the