fix(rules-cuda): the clang route also names the CCCL payload

curand_mtgp32_kernel.h includes <nv/target> from CCCL, which the 12.x
toolkits ship as the separate cuda-cccl package and which the host's
/usr/include had supplied silently as well. The rule adds the payload's
include directory (and include/cccl for the 13.x layout) and refuses the
clang route without it, naming both entries; the fixture declares it.

Author: speak-agent

rules/cuda.cppm

@@ -147,7 +147,10 @@ struct toolkit {
     // cannot compile any device unit without them. Measured 2026-09-05: on a
     // developer machine the header was found in the HOST's /usr/include and
     // the leak went unnoticed until a runner with no host CUDA refused it.
-    std::string curand_root;
+    // The same header then includes <nv/target> from CCCL (libcu++), which the
+    // 12.x toolkits ship as the separate `cuda-cccl` package, and which the
+    // host's /usr/include had supplied in the same way.
+    std::string curand_root, cccl_root;
     std::string nvcc() const { return nvcc_root + "/bin/nvcc"; }
     std::string host_config() const {
         for (auto const* r : { &crt_root, &nvcc_root, &cudart_root }) {
@@ -159,11 +162,19 @@ struct toolkit {
     }
     std::vector<std::string> include_dirs() const {
         std::vector<std::string> out;
-        for (auto const* r : { &cudart_root, &crt_root, &nvcc_root, &curand_root })
+        for (auto const* r : { &cudart_root, &crt_root, &nvcc_root, &cccl_root, &curand_root })
             if (!r->empty() && std::filesystem::is_directory(*r + "/include"))
                 out.push_back(*r + "/include");
+        // The 13.x CCCL payload nests its tree one directory down.
+        if (!cccl_root.empty() && std::filesystem::is_directory(cccl_root + "/include/cccl"))
+            out.push_back(cccl_root + "/include/cccl");
         return out;
     }
+    bool has_cccl() const {
+        return !cccl_root.empty()
+            && (std::filesystem::exists(cccl_root + "/include/nv/target")
+                || std::filesystem::exists(cccl_root + "/include/cccl/nv/target"));
+    }
     std::vector<std::string> lib_dirs() const {
         std::vector<std::string> out;
         for (auto const* r : { &cudart_root, &nvcc_root })
@@ -185,6 +196,7 @@ inline std::optional<toolkit> find_toolkit() {
     t.cudart_root = xpkg("cuda-cudart");
     t.crt_root    = xpkg("cuda-crt");
     t.curand_root = xpkg("libcurand");
+    t.cccl_root   = xpkg("cuda-cccl");
     t.driver_dir  = xpkg("libcuda-host-link");
     if (t.nvcc_root.empty() || t.cudart_root.empty()) {
         std::println(std::cerr,
@@ -444,13 +456,16 @@ inline std::vector<edge> plan(std::span<const std::string> sources, options opt
         // would fail on belongs to a payload the project has to name, and the
         // diagnostic names it. A host copy is never searched for -- that is
         // how the leak above survived every local build.
-        if (tk->curand_root.empty()
-            || !std::filesystem::exists(tk->curand_root + "/include/curand_mtgp32_kernel.h")) {
+        const bool curand_ok = !tk->curand_root.empty()
+            && std::filesystem::exists(tk->curand_root + "/include/curand_mtgp32_kernel.h");
+        if (!curand_ok || !tk->has_cccl()) {
             std::println(std::cerr,
                 "mcpp.rules.cuda: the clang route needs cuRAND's headers, which clang's CUDA "
-                "wrapper includes unconditionally.\n"
-                "  Name the payload under [xlings.workspace] and mcpp provisions it on first use:\n"
-                "    \"xim:libcurand\" = \"10.3.10.19\"   (the 12.9 line; 10.4.x pairs with 13.x)");
+                "wrapper includes unconditionally, and CCCL's, which they include in turn.\n"
+                "  Name the payloads under [xlings.workspace] and mcpp provisions them on first use:\n"
+                "    \"xim:cuda-cccl\" = \"12.9.27\"       (the 12.9 line; 13.x pairs with 13.x)\n"
+                "    \"xim:libcurand\" = \"10.3.10.19\"   (the 12.9 line; 10.4.x pairs with 13.x)\n"
+                "  (found cccl: '{}', curand: '{}')", tk->cccl_root, tk->curand_root);
             return out;
         }
         front = { driver_cc, "-x", "cuda", "-std=c++17", "-O2", "-fPIC",

tests/cuda-consumer/mcpp.toml

@@ -42,6 +42,9 @@ cuda-runtime = "2026.09.05"
 # every device unit, so the clang route needs them even when nothing calls
 # cuRAND. The 10.3.x line pairs with CUDA 12.x.
 "xim:libcurand"         = "10.3.10.19"
+# CCCL (libcu++, cub, thrust): cuRAND's header includes <nv/target> from it.
+# The 12.x toolkits ship it as a separate package; 12.9.27 is the 12.9 line.
+"xim:cuda-cccl"         = "12.9.27"
 "xim:libcuda-host-link" = { linux = "0.0.1" }
 
 [build]