diag: resolver introspection at load

vsa_h3.py

@@ -293,15 +293,23 @@ def _resolve_cuda_sparse_op():
     if _CUDA_SPARSE["resolved"]:
         return _CUDA_SPARSE["op"]
     _CUDA_SPARSE["resolved"] = True
+    if os.environ.get("H3_VSA_CUDA", "1") != "1":
+        return None
+    if _CUDA_SPARSE["resolved"]:
+        return _CUDA_SPARSE["op"]
+    _CUDA_SPARSE["resolved"] = True
     if os.environ.get("H3_VSA_CUDA", "1") != "1":
         return None
     try:
         from fastvideo_kernel import block_sparse_attn_sm100a as sm100a
 
-        op = sm100a.block_sparse_attn_sm100a
-        probe = sm100a._FWD_BY_BLOCK  # noqa: F841 - import-time check that the extension loaded
         if not sm100a._HAS_VSA_SM100A:
-            raise ImportError("fastvideo_kernel compiled without the sm100a VSA forward")
+            # Surface what the extension actually exposes so the logs show which pybind name this wheel has.
+            from fastvideo_kernel._C import fastvideo_kernel_ops as _ops
+
+            available = sorted(a for a in dir(_ops) if "sparse" in a.lower())
+            raise ImportError(f"sm100a forward flag unset; pybind exposes {available}")
+        op = sm100a.block_sparse_attn_sm100a
         _CUDA_SPARSE["op"] = op
         print("[vsa] CUDA sparse-attention kernel: fastvideo_kernel (sm100a fatbin on sm120)", flush=True)
     except Exception as error:  # noqa: BLE001 - any failure means Triton