fix: use explicit zig arch targets instead of native (#21682)

## Summary

Supersedes #21685. Fixes SIGILL crashes on ARM64 Mac devcontainers
caused by zig's `-target native` detecting host CPU features (SVE on
Graviton) and emitting instructions Apple Silicon can't execute.

#21611 tried to fix this with `-march=generic` from cmake, but that's
not a valid AArch64 flag — zig silently ignored it. #21685 fixed ARM
only; this PR also fixes x86 and simplifies the architecture targeting:

- **zig wrappers**: Replace `-target native` with explicit
`aarch64`/`x86_64` targets on both architectures, preventing
host-specific CPU feature detection (SVE, AVX-512, etc.). Error on
unsupported architectures.
- **arch.cmake**: Hardcode `-march=skylake` on x86, skip on ARM. Remove
`TARGET_ARCH` variable and auto-detection logic entirely.
- **CMakePresets.json**: Remove all `TARGET_ARCH` entries from
cross-compile presets (invalid `generic` on arm64-linux, redundant
`skylake` on amd64-linux/windows).

## Note on TARGET_ARCH removal

`TARGET_ARCH` was originally introduced in #2660 (Oct 2023) so the
`publish-bb.yml` GitHub Actions workflow could target `westmere` for
broader x86 compatibility when shipping standalone bb binaries. That
workflow was deleted in #12347 (Feb 2025), so there are no remaining
consumers of configurable arch targeting.

If bb binaries are ever distributed to end users on pre-skylake x86
hardware again, this would need to be reintroduced (e.g. via
`-DCMAKE_CXX_FLAGS=-march=westmere` or a new variable). For now, all
consumers are assumed to be on modern hardware where skylake is the
minimum.

## Test plan
- [ ] CI barretenberg build passes on ARM and x86
- [ ] M3 devcontainer: `NO_CACHE=1` rebuild produces working binaries
(no SIGILL)

Author: AztecBot

barretenberg/cpp/CMakePresets.json

@@ -489,7 +489,6 @@
       },
       "cacheVariables": {
         "CMAKE_SYSTEM_NAME": "Linux",
-        "TARGET_ARCH": "skylake",
         "AVM_TRANSPILER_LIB": "${sourceDir}/../../avm-transpiler/target/x86_64-apple-darwin/release/libavm_transpiler.a"
       }
     },
@@ -503,7 +502,6 @@
       },
       "cacheVariables": {
         "CMAKE_SYSTEM_NAME": "Linux",
-        "TARGET_ARCH": "generic",
         "AVM_TRANSPILER_LIB": "${sourceDir}/../../avm-transpiler/target/aarch64-unknown-linux-gnu/release/libavm_transpiler.a"
       }
     },
@@ -554,7 +552,6 @@
       "cacheVariables": {
         "CMAKE_SYSTEM_NAME": "Windows",
         "CMAKE_SYSTEM_PROCESSOR": "x86_64",
-        "TARGET_ARCH": "skylake",
         "BB_LITE": "ON",
         "AVM_TRANSPILER_LIB": ""
       }

barretenberg/cpp/cmake/arch.cmake

@@ -5,16 +5,9 @@ if(WASM)
     add_compile_options(-fno-exceptions -fno-slp-vectorize)
 endif()
 
-# Auto-detect TARGET_ARCH on x86_64 if not explicitly set (native builds only).
-# On ARM, we skip -march entirely — the zig wrappers use an explicit aarch64 target
-# to produce generic ARM64 code without CPU-specific extensions (e.g. SVE).
-# Skip auto-detection when cross-compiling — the toolchain (e.g. Zig -mcpu) handles
-# architecture targeting, and injecting -march here conflicts with it.
-if(NOT WASM AND NOT TARGET_ARCH AND NOT ARM AND NOT CMAKE_CROSSCOMPILING)
-    set(TARGET_ARCH "skylake")
-endif()
-
-if(NOT WASM AND TARGET_ARCH)
-    message(STATUS "Target architecture: ${TARGET_ARCH}")
-    add_compile_options(-march=${TARGET_ARCH})
+# Target skylake on x86 for AVX2 etc. ARM is handled by the zig wrapper scripts
+# which use explicit aarch64 targets to produce generic ARM64 code without
+# CPU-specific extensions (e.g. SVE on Graviton) that would break on Apple Silicon.
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64")
+    add_compile_options(-march=skylake)
 endif()

barretenberg/cpp/scripts/zig-c++.sh

@@ -1,14 +1,15 @@
 #!/bin/bash
 # Wrapper for zig c++ that pins glibc 2.35 on Linux (Ubuntu 22.04+ compat)
 # and uses native target on macOS.
-# On ARM64 Linux, use an explicit aarch64 target instead of 'native' to produce
-# generic ARM64 code. This prevents CPU-specific instructions (e.g. SVE on Graviton)
-# from being emitted, ensuring binaries work across all ARM64 machines including
-# Apple Silicon in devcontainers.
+# Use explicit architecture targets instead of 'native' to prevent zig from
+# detecting host-specific CPU features (e.g. SVE on Graviton, AVX-512 on
+# Sapphire Rapids) that would produce binaries incompatible with other machines.
+# cmake's arch.cmake handles -march=skylake for x86; ARM gets baseline aarch64 (no SVE).
 if [[ "$(uname -s)" == "Linux" ]]; then
   case "$(uname -m)" in
     aarch64|arm64) exec zig c++ -target aarch64-linux-gnu.2.35 "$@" ;;
-    *)             exec zig c++ -target native-linux-gnu.2.35 "$@" ;;
+    x86_64|amd64)  exec zig c++ -target x86_64-linux-gnu.2.35 "$@" ;;
+    *)             echo "Error: unsupported architecture '$(uname -m)'" >&2; exit 1 ;;
   esac
 else
   exec zig c++ "$@"

barretenberg/cpp/scripts/zig-cc.sh

@@ -1,14 +1,15 @@
 #!/bin/bash
 # Wrapper for zig cc that pins glibc 2.35 on Linux (Ubuntu 22.04+ compat)
 # and uses native target on macOS.
-# On ARM64 Linux, use an explicit aarch64 target instead of 'native' to produce
-# generic ARM64 code. This prevents CPU-specific instructions (e.g. SVE on Graviton)
-# from being emitted, ensuring binaries work across all ARM64 machines including
-# Apple Silicon in devcontainers.
+# Use explicit architecture targets instead of 'native' to prevent zig from
+# detecting host-specific CPU features (e.g. SVE on Graviton, AVX-512 on
+# Sapphire Rapids) that would produce binaries incompatible with other machines.
+# cmake's arch.cmake handles -march=skylake for x86; ARM gets baseline aarch64 (no SVE).
 if [[ "$(uname -s)" == "Linux" ]]; then
   case "$(uname -m)" in
     aarch64|arm64) exec zig cc -target aarch64-linux-gnu.2.35 "$@" ;;
-    *)             exec zig cc -target native-linux-gnu.2.35 "$@" ;;
+    x86_64|amd64)  exec zig cc -target x86_64-linux-gnu.2.35 "$@" ;;
+    *)             echo "Error: unsupported architecture '$(uname -m)'" >&2; exit 1 ;;
   esac
 else
   exec zig cc "$@"