[Feature Request] Architecture-optimized native builds (PGO/BOLT/CPU-specific)

Summary

Request for optimized native builds targeting modern CPU architectures across all supported platforms.

Current State

• Native installer downloads generic baseline binaries
• No CPU-specific optimizations (no AVX-512, no Zen tuning, no Apple Silicon optimization)
• No PGO (Profile-Guided Optimization) applied
• No BOLT (Binary Optimization and Layout Tool) post-processing on Linux

Requested Options

Architecture Variants

| Platform | Current | Requested Variants |
|----------|---------|-------------------|
| Linux x86-64 | GNU/Linux 3.2.0 baseline | znver4/znver5 (AMD Zen 4/5), alderlake/sapphirerapids (Intel 12th+) |
| Linux ARM64 | Generic ARMv8 | neoverse-v2, cortex-a76 (AWS Graviton, Ampere) |
| macOS ARM64 | Generic Apple Silicon | apple-m1, apple-m2, apple-m3 (native Apple tuning) |
| macOS x86-64 | Generic x86-64 | haswell, skylake (Intel Mac baseline) |
| Windows x86-64 | Generic x86-64 | znver4, alderlake (matching Linux variants) |

Optimization Techniques

1. PGO builds: Pre-profiled binaries with representative Claude workloads
2. BOLT-optimized builds (Linux): Post-link optimization for better I-cache utilization
3. LTO (Link-Time Optimization): Cross-module optimization for smaller, faster binaries

Installer Enhancements

# Examples of desired installer flags
claude-installer --arch=native          # Auto-detect and use best variant
claude-installer --arch=znver5          # Explicit architecture selection
claude-installer --optimize=pgo,bolt    # Select optimization level

Expected Benefits

| Optimization | Expected Gain | Applies To |
|--------------|---------------|------------|
| Architecture-specific codegen | 5-15% | All platforms |
| PGO (Profile-Guided) | 10-20% | All platforms |
| BOLT (hot code reordering) | 5-15% | Linux only |
| LTO | 5-10% binary size, 2-5% perf | All platforms |

Combined: 15-30% faster startup and local operations

Technical Notes

Bun Single-File Executable Considerations

Claude Code uses Bun's single-file executable format, which appends the JavaScript bundle after the ELF/Mach-O structure. Optimizations must be applied before bundling:

1. Build Bun with PGO using Claude-representative workloads
2. Apply BOLT to Bun runtime before JS payload append
3. Use architecture-specific compiler flags during Bun compilation

Why Users Can't Self-Optimize

• BOLT strips the appended JS bundle, rendering the binary non-functional
• PGO requires source access and representative workload profiling
• Architecture flags must be set at compile time

Environment (Reporter)

• CPU: AMD Ryzen 7 9800X3D (Zen 5, AVX-512)
• OS: Arch Linux 6.18
• Claude Code: 2.1.29

Related

• Bun single-file executables
• LLVM BOLT
• PGO in LLVM