build: C++26 by default (amdclang 23 / g++ 15), no module scanning

[bong-water-water-bong]

The fork builds at C++26 by default, like the engine (CMAKE_CXX_STANDARD 26 in the top level and ggml/, CMAKE_CXX_STANDARD_REQUIRED kept, CMAKE_CXX_SCAN_FOR_MODULES OFF since TheRock has no clang-scan-deps). Per-target cxx_std_17 lines stay (they are minimums); HIP code objects keep their explicit -std=c++17. 16 files, +95/−36, upstream-style edits, no new files.

Source fixes the standard needed

• src/models/models.h: declares the template <> specializations of the eagle3/dflash/t5 graph<...> constructors (and build_inp_embd_enc) after each struct. Since C++23 libstdc++'s make_unique is constexpr, so clang instantiates it inside build_arch_graph before the specialization was declared — the code was ill-formed NDR in every standard; this also fixes t5encoder.cpp.
• common/chat.cpp: std::optional<json>(std::in_place, …) — the json → optional<json> conversion is ambiguous at C++26 with libstdc++ 16.
• enum | enum of different types is ill-formed since C++26: casts in tools/mtmd/clip-graph.h, tools/mtmd/models/qwen3vl.cpp, tests/test-backend-ops.cpp (6 sites incl. var_to_str).
• ggml/src/ggml-hrx/runtime/kernel-executable-cache.cpp: std::atomic<std::shared_ptr<…>> instead of the deprecated atomic_load_explicit on a shared_ptr.
• ggml/src/ggml-backend-reg.cpp: path_from_u8() (from std::u8string when __cpp_lib_char8_t) replaces the 8 deprecated fs::u8path calls.
• vendor/nlohmann/json.hpp:17557: is_trivial (deprecated) → is_trivially_default_constructible && is_trivially_copyable (removes 43 g++ warnings).

Silent behaviour change caught and neutralised. In C++26 std::to_string(float/double) formats like std::format("{}") instead of %f. It compiled, but test-jinja failed 5 cases ({{ -1.0 }} → -1, 42|float → 42; through the jinja trailing-zero trim 100.0 would print 1 and 0.0 an empty string followed by out.back() on it). All 17 call sites in 7 files now print %f as before (common/jinja/{string,value}.h, src/llama-impl.cpp, tools/mtmd/clip-impl.h, tools/llama-bench/llama-bench.cpp, tools/server/server-schema.cpp, tests/test-backend-ops.cpp), found by a temporary [[deprecated]] to_string(float) overload compiled through every TU with both compilers.

Checks (Strix Halo, performance mode; amdclang 23.0.0git TheRock + libstdc++ 16, g++ 15.2.0, cmake 4.2.3)

• 0 warnings from fork sources on both compilers (same as the C++17 baseline).
• test-backend-ops -b HRX0 ×2, C++17 and C++26 builds identical: MUL_MAT 325, MUL_MAT_ID 108, MUL_MAT_VEC_FUSION 57, ROPE 42, RMS_NORM 6, SOFT_MAX 8, CONCAT 6, SSM_CONV 47.
• KLD Qwen3-0.6B c512 ×16: the saved logits of the C++17 and C++26 builds are byte-identical (PPL 26.1766 both).
• 8 identical greedy requests: 1/8 distinct on Qwen3-0.6B and ZAYA1-8B, byte-identical to the C++17 build.
• ctest (g++ CPU): same results at 17 and 26; test-jinja passes.
• llama-bench HRX0, pp512/tg128 separate processes, 3 interleaved rounds × -r 3, medians of 9: Qwen3-0.6B −0.31% / −0.08%, Qwen3-8B +0.42% / +0.11%, GLM-4.7-Flash −0.12% / +0.05%, ZAYA1-8B −0.05% / +0.15% — all within noise.

Not in this PR: the private add-on's two GLM HIP harnesses needed C++23 (clang 23 cuda_wrappers/new vs libstdc++ 16 constexpr placement delete) — fixed in that repo. The -MD -MF on the fork's --cuda-device-only code-object rule gives 35 unused-argument warnings and probably never writes the depfiles — separate follow-up.

🤖 Generated with Claude Code