fix(build): pass -Wno-template-body to nvcc host compiler for gcc 15

GCC 15 (Fedora 43, the cuda-13.0 runner base) diagnoses errors in
uninstantiated template bodies, tripping over a latent typo
(set_slice3x3) in the CUTLASS matrix.h pulled in by mistralrs-quant's
cutlass_moe kernels. The offending Matrix::rotation/reflection members
are never instantiated, so defer the diagnosis via NVCC_APPEND_FLAGS
(read directly by nvcc) instead of failing every build/flavour.

Also reduce the upstream poll cron from 6h to 1h for quicker feedback;
the in-flight guard already prevents re-dispatch over a running build.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

.gitea/workflows/build-prerelease.yml

@@ -84,6 +84,13 @@ jobs:
           CUDA_COMPUTE_CAP: ${{ matrix.compute_caps }}
           CARGO_BUILD_JOBS: ${{ matrix.build_jobs }}
           NVCC_THREADS: ${{ matrix.nvcc_threads }}
+          # GCC 15 (Fedora 43) diagnoses errors in uninstantiated template
+          # bodies, tripping over a latent typo (set_slice3x3) in the CUTLASS
+          # matrix.h pulled in by mistralrs-quant's cutlass_moe kernels. The
+          # offending Matrix::rotation/reflection members are never instantiated,
+          # so defer the diagnosis instead of failing the build. nvcc reads
+          # NVCC_APPEND_FLAGS directly and forwards -Wno-template-body to gcc.
+          NVCC_APPEND_FLAGS: "-Xcompiler -Wno-template-body"
 
       - name: Collect artifacts
         run: |

.gitea/workflows/build-release.yml

@@ -71,6 +71,13 @@ jobs:
           CUDA_COMPUTE_CAP: ${{ matrix.compute_caps }}
           CARGO_BUILD_JOBS: ${{ matrix.build_jobs }}
           NVCC_THREADS: ${{ matrix.nvcc_threads }}
+          # GCC 15 (Fedora 43) diagnoses errors in uninstantiated template
+          # bodies, tripping over a latent typo (set_slice3x3) in the CUTLASS
+          # matrix.h pulled in by mistralrs-quant's cutlass_moe kernels. The
+          # offending Matrix::rotation/reflection members are never instantiated,
+          # so defer the diagnosis instead of failing the build. nvcc reads
+          # NVCC_APPEND_FLAGS directly and forwards -Wno-template-body to gcc.
+          NVCC_APPEND_FLAGS: "-Xcompiler -Wno-template-body"
 
       - name: Collect artifacts
         run: |

.gitea/workflows/poll-upstream.yml

@@ -2,7 +2,7 @@ name: poll-upstream
 
 on:
   schedule:
-    - cron: "0 */6 * * *"
+    - cron: "0 * * * *"
   workflow_dispatch: {}
 
 concurrency: