[jtuyls]

Reduce the packaged compiler payload without changing its public tool interfaces.

• Compress compiler wheels with DEFLATE level 9.
• Replace byte-identical Windows Clang, LLD, and readelf multicall copies with deterministic native launchers that preserve each driver flavor.
• Replace the large Linux LLD and readelf copies with flavor-preserving shell launchers while retaining the inexpensive Clang aliases.
• Remove the byte-identical Linux libLLVM-21.so copy after validation and strip unneeded ELF symbols.
• Exercise every forwarded mode in the existing wheel verification jobs and retain native dependency, installed-wheel, E2E, and JIT checks.

Measured compiler-core wheel sizes:

• Windows: 632,790,558 -> 330,049,707 bytes (47.8% smaller).
• Linux: 469,932,885 -> 303,792,202 bytes (35.4% smaller).

[Copilot]

🟡 Changes recommended

Launchers regress direct Windows Clang frontend calls and Linux alias invocation through external symlinks.

2 open issues

• Resolve launcher symlinks before locating sibling executables · New
• Avoid injecting driver mode for direct Clang frontend calls · New

What changed in this PR

Reduces compiler wheel size through stronger compression and payload deduplication while aiming to preserve existing tool interfaces.

Changes:

• Uses DEFLATE level 9 and replaces duplicate tool binaries with launchers.
• Removes redundant Linux LLVM payload and strips unneeded symbols.
• Expands alias verification and documents packaging behavior.

File	Description
tests/​wheels/​test_fastflow_sdk_compiler_wheels.py	Tests alias forwarding, validation, and stripping.
docs/​wheel-releases.md	Documents size-reduction and verification behavior.
build_tools/​native/​compiler_tool_alias_launcher.cpp	Adds native Windows alias forwarding.
build_tools/​fastflow_sdk_compiler_wheels.py	Implements compression, payload reduction, and alias audits.

_{Review effort: Balanced}