fix(train): accept wandb_v1_ key format in W&B gate check

isalnum() rejects the new wandb_v1_XXX key format which contains
underscores. Widen the check to allow underscores so W&B logging
is not silently disabled for users with new-format API keys.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

train/grpo.py

@@ -135,7 +135,8 @@ def run_grpo(
             "near 1.0 (avoids QLoRA + vLLM merge drift bug)."
         )
     wb_key = os.environ.get("WANDB_API_KEY", "").strip()
-    use_wandb = len(wb_key) >= 40 and wb_key.isalnum()
+    # Accept both old-style (40-char alnum) and new wandb_v1_ keys (contain underscores).
+    use_wandb = len(wb_key) >= 40 and all(c.isalnum() or c == "_" for c in wb_key)
     _set_if_supported(["report_to"], ["wandb"] if use_wandb else [])
     if not use_wandb:
         log.warning(