fix(interface): corex_moe accepts w13 merged format + no silent fallback

corex_moe.py: moe_forward now accepts both formats:
  Format A: w1(E,I,H) + w2(E,H,I) + w3(E,I,H) — xllm style, separate gate/up
  Format B: w13(E,2*I,H) + w2(E,H,I) + w3=None — vllm style, merged gate_up
  Auto-detects by checking if w3 is None, splits w13 internally.

qwen3_5.py:
  - Fix corex_moe call: use keyword args (w3=None, topk=self.top_k)
    prevents topk integer going to w3 tensor position
  - Remove silent fallback on corex_moe failure — raise RuntimeError
    with full shape info for diagnosis. Zero score with no error log
    is worse than a crash.

ex_engine/python/corex_moe.py

@@ -127,20 +127,34 @@ def topk_softmax(
 def moe_forward(
     hidden_states: torch.Tensor,
     gate_output: torch.Tensor,
-    w1: torch.Tensor,
+    w1_or_w13: torch.Tensor,
     w2: torch.Tensor,
-    w3: torch.Tensor,
+    w3: Optional[torch.Tensor] = None,
     topk: int = 8,
     renormalize: bool = True,
     **kwargs,
 ) -> torch.Tensor:
     """
     Full MoE pipeline: CUDA topk → per-expert GEMM (cublas) → silu → GEMM → scatter-add.
+
+    Accepts two weight formats:
+      Format A (xllm style):  w1=(E,I,H), w2=(E,H,I), w3=(E,I,H) — gate and up separate
+      Format B (vllm style):  w13=(E,2*I,H), w2=(E,H,I), w3=None — gate_up merged
     """
     num_tokens = hidden_states.shape[0]
     hidden_size = hidden_states.shape[1]
     dtype = hidden_states.dtype
 
+    # Detect weight format
+    if w3 is None:
+        # Format B: w13 merged — split into w1 (gate) and w3 (up)
+        w13 = w1_or_w13
+        inter2 = w13.shape[1]
+        w1 = w13[:, :inter2 // 2, :]  # (E, I, H)
+        w3 = w13[:, inter2 // 2:, :]  # (E, I, H)
+    else:
+        w1 = w1_or_w13
+
     topk_weights, topk_ids = topk_softmax(gate_output, topk, renormalize)
 
     num_experts = w1.shape[0]