feat(JuliaStrings/utf8proc): add utf8proc v2.9.0 LLAR formula

[fennoai]

Closes #22.

Translates the Conan Center utf8proc recipe (snapshot ffe30df1, version 2.9.0) into an idiomatic LLAR Formula for JuliaStrings/utf8proc.

Files

• JuliaStrings/utf8proc/versions.json — no dependencies (utf8proc is a self-contained single-file C library; upstream v2.9.0 declares no third-party requirements).
• JuliaStrings/utf8proc/Utf8proc_cmp.gox — semver comparator for the leading-v release tags.
• JuliaStrings/utf8proc/v2.9.0/Utf8proc_llar.gox:
  • onBuild uses the CMake helper to build the static archive (the upstream and Conan default). Pins CMAKE_POLICY_VERSION_MINIMUM=3.5 and CMAKE_POLICY_DEFAULT_CMP0042=NEW exactly as the recipe does for pre-2.10.0 releases, CMAKE_POSITION_INDEPENDENT_CODE=ON to match Conan's default fPIC=True, and UTF8PROC_ENABLE_TESTING=OFF to keep the build hermetic.
  • Metadata is derived from the installed libutf8proc.pc via pkg-config, plus -DUTF8PROC_STATIC to mirror the Conan static package_info branch.
  • onTest compiles and runs a consumer modelled on the Conan test_package (utf8proc_version / utf8proc_unicode_version) in a build tree independent of the onBuild scratch dir, so it also passes on a cache hit.

Version boundary

fromVer is pinned to v2.9.0. The earlier versions served by the same Conan folder (e.g. v2.5.0) use a lower CMake floor (2.8.12) and do not install a pkg-config file, so their build contract differs and is not verified here. A lower threshold should be added only after those revisions are inspected and tested — per the translation guidance to add a threshold rather than claim an unverified range.

Notes on options

The Conan shared/fPIC options were reviewed. In the current llar, matrix --option overrides propagate to the cache key/install path but are not injected into the formula body (target.options only ever reflects defaults, and there is no public accessor for the matrix string). Rather than expose a shared toggle that would silently always build static, this formula builds the static default (matching Conan shared=False) and documents fPIC=True behavior directly. A shared option can be added once option selection reaches onBuild.

Validation

Built llar from source and ran on linux/amd64:

llar test -v ./JuliaStrings/utf8proc@v2.9.0     # cache-miss: PASS
llar test    ./JuliaStrings/utf8proc@v2.9.0     # cache-hit (onBuild skipped, onTest runs): PASS
llar test --os linux --arch amd64 ./JuliaStrings/utf8proc@v2.9.0   # PASS
llar make --json ./JuliaStrings/utf8proc@v2.9.0

The install produces include/utf8proc.h, lib/libutf8proc.a, and lib/pkgconfig/libutf8proc.pc; the consumer prints utf8proc version: 2.9.0 and Unicode version: 15.1.0. No system libutf8proc is present, confirming the test links the freshly built artifact.

[fennoai]

Review summary

This PR adds a new JuliaStrings/utf8proc formula (build + test + semver comparator) and refactors madler/zlib onto the current single-context Formula API. The zlib refactor, the _cmp.gox, versions.json, and the large docs/skill extraction all look correct and internally consistent (mirrored .claude/.agents copies match, internal links resolve, docs describe the new API rather than the removed error-return surface).

One important correctness issue in the utf8proc onBuild metadata step, plus one minor robustness note — both inline.

Non-blocking observations (no change required):

• onTest recompiles the C consumer on every run including cache hits. This is intentional per the inline comment (cache-hit/cache-miss parity) and the compile is tiny; noted only for awareness.
• SKILL.md frontmatter differs between the .claude (has argument-hint/allowed-tools) and .agents copies. Likely deliberate since .agents uses agents/openai.yaml for interface metadata — worth a quick confirm.

Findings without inline locations

• JuliaStrings/utf8proc/v2.9.0/Utf8proc_llar.gox:86: Minor robustness: this line uses sh -c only to run a fixed binary with no shell features. Invoking it directly with exec "'" ... is unnecessary — exec bin would run the binary without a shell round-trip and sidesteps the single-quote escaping concern entirely (single-quoting protects against a | in the matrix suffix but not an embedded ' in the path). Not exploitable given these are trusted sandbox paths; purely a cleanliness/robustness note.

[fennoai]

Missing lastErr check before reading output. After a capout block, the reference convention requires validating the command result (if lastErr != nil { panic lastErr }) before consuming output — the onTest sections below do exactly this, and so does the reference's complete example.

Here, if pkg-config fails (missing .pc, wrong name, or not installed), the hook does not error: output is empty/stale, so metadata is silently set to just -DUTF8PROC_STATIC, producing an install with broken consumer link/compile flags instead of a clear build failure.

capout => {
    exec "pkg-config", "--libs", "--cflags", "libutf8proc"
}
if lastErr != nil {
    panic lastErr
}
meta := strings.trimSpace(output)