azazel-parity-ghostty

ghostty's UTF8Decoder — a self-contained, std-only terminal UTF-8 decoder — built two ways, to prove and compare azazel and zaza.

Both build a consumer that decodes "a" + U+4E16 (世) byte by byte and checks the codepoints. Neither vendors the source; each stages the single decoder file at a pinned commit with its own fetch.sh (a curl, no clone).

Pinned upstream

Repository	https://github.com/ghostty-org/ghostty
Commit	168c7b94672d91cded4b506143cb0ebebc5d1ceb
File	src/terminal/UTF8Decoder.zig
Zig	0.16.0

Build it

cd azazel && ./fetch.sh && sh gen_build_spec.sh && zig build && ./zig-out/bin/consumer
cd zaza  && ./fetch.sh && zig build run

Both print 2 codepoints (97, 0x4e16): a (97) and 世 (U+4E16).

Comparison

Clean-cache builds with dependencies pre-fetched, Apple Silicon, fastest of two runs.

Build	Clean build	Config
azazel	4.6 s	project.cue — 14 lines · 455 B
zaza	4.4 s	build.zig — 30 lines · 1059 B

The upstream's full build is not reproduced here (see the note below), so no native time is listed.

azazel and zaza land within ~0.2 s of each other; ghostty's full build needs system libraries and is not reproduced here.

Build process & what can be optimized

Both build roots stage the pinned upstream with fetch.sh into a git-ignored vendor/ (a curl for single-file slices, a shallow clone for source trees) — no upstream sources are committed. Then:

• azazel: sh gen_build_spec.sh runs CUE and emits build_spec.zig (the build declared as data), then zig build compiles it. The CUE step is memoized — it re-runs only when the model changes (~0.20s → ~0.01s otherwise).
• zaza: zig build drives the standard Zig build graph directly.

What actually makes it faster

Measured across the corpus (clean vs warm builds):

Lever	Speedup	Note
Content-addressed cache (rebuild)	89×	14.2s → 0.16s; Zig has it, both inherit it
Incremental (edit one file)	10.8×	14.2s → 1.32s; deps stay cached
CI dependency cache	2×	cold 13.3s → warm 6.6s; this repo's CI caches ~/.cache/zig
Memoized CUE codegen	20×	azazel's only overhead, gone
Parallelism (many cores)	1.1×	marginal — shared std + startup dominate
GPU	none	compilation is branchy, sequential, dependency-ordered

The instinct to parallelize like a C++ build doesn't transfer: Zig is one mostly-single-threaded compile per artifact with a fast self-hosted backend and a shared std that caches. For Zig, caching is the lever, not parallelism.

The real frontier is residency: a resident compile server that keeps the InternPool hot and recompiles only changed declarations, plus in-place binary patching (Zig's roadmap) and a shared content-addressed cache. azazel's build-as-data is positioned for it — the build is a query, and the cache key is computable from the pinned model without running the compiler. Full write-up and the cross-repo comparison: the corpus dashboard.