cmake: wire up fastcache-cc / sccache / ccache compiler-cache selection

[Yaraslaut]

Vendors cmake/CompileCache.cmake from LASTRADA-Software/fastcached — a self-contained module with no dependency on that project, meant to be dropped into any CMake project's cmake/ directory (see cmake/portable/README.md there).

What it does

Selects a compiler-cache launcher in preference order, wired via CMAKE_<LANG>_COMPILER_LAUNCHER:

1. fastcache-cc — used only when a fastcached daemon actually answers. This is verified end-to-end at configure time by compiling a throwaway translation unit through the launcher and requiring a reported HIT/MISS, not just presence on PATH (a launcher that can't reach its daemon still compiles fine, silently paying a failed connect per TU otherwise). Address defaults to 127.0.0.1:6674, overridable via FASTCACHE_ADDR.
2. sccache — used when fastcache-cc is unavailable/unconfigured.
3. ccache — the local fallback.
4. None — silently disables caching if nothing above applies, or if -DUSE_COMPILER_CACHE=OFF.

include(cmake/CompileCache.cmake) is placed right after compiler_options.cmake, before the FetchContent-fetched glaze dependency, so that gets cached too.

Why fastcache-cc is preferred over sccache

fastcache-cc isn't just "another sccache" — when it's selected, the compile cache is served over fastcached's own dedicated 0xFC binary wire protocol (CompileCacheWire/CompileCacheHandler), purpose-built for this job: canonicalize-on-store / serve-canonical-on-fetch, cohort-based leading-key prefetch, and a framing that lets a rejection be a reply instead of a silent connection drop. sccache talks a generic key/value protocol against whatever backend (Redis, S3, ...) it's pointed at — it works, but it isn't using our protocol. We want our own protocol used whenever a fastcached daemon is available, and this preference order is how that's expressed: try fastcache-cc first, fall back only when it can't be used.

Both launchers are portable across checkout paths (unlike raw ccache in some configurations), so a CI runner and a developer working from different directories can share cache hits. This has no effect on a machine with nothing installed — it's a pure opt-in speedup when a cache is available.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!