ci: strip Blackwell sm_100 too (5.6.1 arch matrix broke the runner)

PhysX 5.6.1 added --generate-code for compute_100 (Blackwell), which the pinned
CUDA 12.5 toolkit cannot compile ("nvcc fatal: Unsupported gpu architecture
'compute_100'"), failing every Windows GPU job. The arch-strip sed only removed
sm_80/86/89/90.

Widen the strip to every non-75 arch: compute_8[0-9] / 9[0-9] / 1[0-9][0-9].
sm_75 stays as the sole real target, PTX JIT covers Ampere..Blackwell at launch.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: Lumengine

.github/workflows/BuildPhysX.yml

@@ -63,20 +63,23 @@ jobs:
         working-directory: physx
         run: |
           # The 16 GB GitHub Windows runner cannot fit the full PhysX arch matrix:
-          # PhysX 5.6 emits --generate-code for sm_75/80/86/89/90 plus `-t0`, which
-          # spawns one ptxas per arch in parallel and crashes ptxas with
-          # ACCESS_VIOLATION (memory exhaustion). PX_GENERATE_GPU_REDUCED_ARCHITECTURES
-          # only drops sm_70, not enough.
+          # PhysX 5.6 emits --generate-code for sm_75/80/86/89/90 (and 5.6.1 adds
+          # Blackwell sm_100) plus `-t0`, which spawns one ptxas per arch in parallel
+          # and crashes ptxas with ACCESS_VIOLATION (memory exhaustion). On top of
+          # that, sm_100 (compute_100) is not even supported by the CUDA 12.5 toolkit
+          # pinned above ("nvcc fatal: Unsupported gpu architecture 'compute_100'").
+          # PX_GENERATE_GPU_REDUCED_ARCHITECTURES only drops sm_70, not enough.
           #
-          # Strategy: keep sm_75 only and rely on PTX JIT to cover Ampere/Ada/Hopper
-          # at first launch, then force nvcc to serialize ptxas per .cu file (-t1).
+          # Strategy: keep sm_75 only and rely on PTX JIT to cover Ampere/Ada/Hopper/
+          # Blackwell at first launch, then force nvcc to serialize ptxas per .cu (-t1).
           shopt -s globstar nullglob
           changed=0
           for f in compiler/vc17win64-gpu-md/**/*.vcxproj; do
-            # Strip sm_80 / sm_86 / sm_89 / sm_90 generate-code entries.
+            # Strip EVERY generate-code entry that is not sm_75: compute_80-99 and
+            # compute_100+ (Blackwell/newer). compute_75 (7x) is left untouched.
             # NB: `#` delimiter avoids clashing with `|` alternation inside the
             # pattern (which `s|...|` would mis-parse as the end of the regex).
-            sed -i -E 's# --generate-code=arch=compute_(80|86|89|90),code=\[[^]]+\]##g' "$f"
+            sed -i -E 's# --generate-code=arch=compute_(8[0-9]|9[0-9]|1[0-9][0-9]),code=\[[^]]+\]##g' "$f"
             # Serialize ptxas threads per nvcc invocation.
             sed -i 's| -t0 | -t1 |g' "$f"
             changed=$((changed+1))