fix(v4): device-aware Path-C status reason (CUDA EAGER bridge vs Metal DSL)

Mirrors the Path-B status fix (3ef4894). _path_c_status() for GDN/KDA reported a
hardcoded "tilelang.compile(target='metal')" even on a CUDA host where Path-C
actually routes through the host _cuda_eager bridge (target='cuda'). Now uses the
path_c module's _device_can_run_metal() to report truthfully (RULE #1: honest
reporting). Metal host text unchanged (verified non-regression).

Author: apstenku123

cppmega_v4/_tilelang/kda_paths.py

@@ -141,19 +141,31 @@ def _path_b_call(*args, allow_fallback: bool = True, **kwargs):
 
 def _path_c_status() -> PathStatus:
     try:
-        from cppmega_v4._tilelang.kda_path_c import _path_c_runtime_status
+        from cppmega_v4._tilelang.kda_path_c import (
+            _path_c_runtime_status,
+            _device_can_run_metal,
+        )
     except Exception as exc:
         return PathStatus(
             path="path_c", available=False,
             reason=f"path_c module not importable: {exc}",
         )
     ok, reason = _path_c_runtime_status()
+    # Device-aware (mirrors _path_b_status): Apple compiles target='metal';
+    # CUDA host routes the same recurrence via the host _cuda_eager bridge.
+    if _device_can_run_metal():
+        detail = (
+            "KDA Path C: TileLang DSL @T.prim_func → tilelang.compile("
+            "target='metal', execution_backend='tvm_ffi')."
+        )
+    else:
+        detail = (
+            "KDA Path C via TileLang-CUDA EAGER bridge (kda_fwd_cuda_eager; "
+            "target='cuda', Metal unavailable on this CUDA host)."
+        )
     return PathStatus(
         path="path_c", available=ok,
-        reason=(
-            f"KDA Path C: TileLang DSL @T.prim_func → tilelang.compile("
-            f"target='metal', execution_backend='tvm_ffi'). {reason}"
-        ),
+        reason=f"{detail} {reason}",
     )
 
 

cppmega_v4/_tilelang/linear_attention_paths.py

@@ -161,19 +161,30 @@ def _path_c_status() -> PathStatus:
     try:
         from cppmega_v4._tilelang.linear_attention_path_c import (
             _path_c_runtime_status,
+            _device_can_run_metal,
         )
     except Exception as exc:
         return PathStatus(
             path="path_c", available=False,
             reason=f"path_c module not importable: {exc}",
         )
     ok, reason = _path_c_runtime_status()
+    # Device-aware (mirrors _path_b_status): on Apple Path C compiles the
+    # TileLang DSL for target='metal' (tvm_ffi); on a CUDA host it routes the
+    # same recurrence through the host _cuda_eager bridge (target='cuda').
+    if _device_can_run_metal():
+        detail = (
+            "GDN Path C: TileLang DSL @T.prim_func → tilelang.compile("
+            "target='metal', execution_backend='tvm_ffi')."
+        )
+    else:
+        detail = (
+            "GDN Path C via TileLang-CUDA EAGER bridge (gdn_fwd_cuda_eager; "
+            "target='cuda', Metal unavailable on this CUDA host)."
+        )
     return PathStatus(
         path="path_c", available=ok,
-        reason=(
-            f"GDN Path C: TileLang DSL @T.prim_func → tilelang.compile("
-            f"target='metal', execution_backend='tvm_ffi'). {reason}"
-        ),
+        reason=f"{detail} {reason}",
     )