Fix nvcc CUDA 12.0/12.1 C++20 ranges parse error (#3907) (#5248)

* Test ci_cuda_example against a CUDA version matrix at C++20 (#3907)

The ci_cuda_example job compiled against the json-ci image's CUDA
11.0 toolkit at cuda_std_11, which cannot exercise #3907 (a c++20
parse error in iteration_proxy.hpp's enable_borrowed_range reported
under nvcc). Switch the job to pull official nvidia/cuda devel images
directly and matrix across CUDA 11.8-12.6 at cuda_std_20 so CI can
empirically confirm which versions are actually affected before any
source-level fix is attempted.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Fix nvcc CUDA 12.0/12.1 C++20 ranges parse error (#3907)

The diagnostic matrix in this PR confirmed the affected range exactly:
nvcc 12.0.1 and 12.1.1 both fail with "expected initializer before
'<' token" on iteration_proxy.hpp's enable_borrowed_range variable
template specialization at -std=c++20; 12.2.2 and newer already build
cleanly. Guard JSON_HAS_RANGES off for that narrow nvcc version range,
matching the existing GCC-11/libstdc++ carve-outs in the same ifdef
chain, and regenerate single_include accordingly.

Broaden the CUDA smoke test to also exercise comparisons
(operator==/operator<=>, gated independently by
JSON_HAS_THREE_WAY_COMPARISON) and range-based iteration, not just
dump()/erase(), so the fix's actual scope is evidenced by CI rather
than assumed from the single reported symptom.

Have tests/cuda_example/CMakeLists.txt pick the newest C++ standard
the detected nvcc version actually supports (20/17/11) instead of
hard-requiring C++20, so older toolkits build at a lower standard
instead of failing CMake configure outright. This is test-project-local
only; the JSON_HAS_RANGES guard is what protects real client code,
since a header can't control what -std= flag it's compiled with.

Right-size the CI matrix from the 8-version diagnostic sweep down to
11.8.0 (C++17 fallback path) / 12.1.1 (permanent #3907 regression
guard) / 12.6.3 (recent coverage), and update the compiler-version
table in the quality assurance docs to match.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Fix ci_cuda_example CUDA 11.8 build after C++17 fallback (#3907)

The 11.8.0 leg's graceful C++17 fallback (added in the previous commit)
worked correctly, but the broadened smoke test used the <=> operator
unconditionally, which isn't valid syntax pre-C++20 — nvcc rejected it
with "expected an expression" once the CMake logic picked cuda_std_17
for the older toolkit. Gate those two lines behind
JSON_HAS_THREE_WAY_COMPARISON like the library itself does internally.

Sanity-compiled the file as plain C++ at both -std=c++17 (skips the
guarded block) and -std=c++20 (includes it) locally; the actual nvcc
build is verified via CI on PR #5248.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

---------

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

tests/cuda_example/CMakeLists.txt

@@ -3,7 +3,18 @@ project(json_cuda LANGUAGES CUDA)
 
 add_executable(json_cuda json_cuda.cu)
 target_include_directories(json_cuda PRIVATE ../../include)
-target_compile_features(json_cuda PUBLIC cuda_std_11)
+
+# nvcc added C++20 support in CUDA 12.0 and C++17 in CUDA 11.0; pick the
+# newest standard the detected compiler actually supports (see #3907)
+# instead of hard-requiring one standard for every CUDA version.
+if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.0)
+    set(json_cuda_std 20)
+elseif(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 11.0)
+    set(json_cuda_std 17)
+else()
+    set(json_cuda_std 11)
+endif()
+target_compile_features(json_cuda PUBLIC cuda_std_${json_cuda_std})
 set_target_properties(json_cuda PROPERTIES
     CUDA_EXTENSIONS OFF
     CUDA_STANDARD_REQUIRED ON

tests/cuda_example/json_cuda.cu

@@ -16,4 +16,20 @@ int main()
     // regression for #3013 (ordered_json::reset() compile error with nvcc)
     nlohmann::ordered_json metadata;
     metadata.erase("key");
+
+    // exercise comparisons (operator==/operator<=>, gated by
+    // JSON_HAS_THREE_WAY_COMPARISON, independent of JSON_HAS_RANGES) and
+    // range-based iteration (exercises iteration_proxy/ranges machinery
+    // beyond just the enable_borrowed_range specialization) — see #3907
+    nlohmann::json a = {1, 2, 3};
+    nlohmann::json b = {1, 2, 3};
+    static_cast<void>(a == b);
+#if JSON_HAS_THREE_WAY_COMPARISON
+    static_cast<void>(a <=> b); // *NOPAD*
+    static_cast<void>(a <=> 1); // *NOPAD*
+#endif
+    for (const auto& element : a)
+    {
+        static_cast<void>(element);
+    }
 }