refactor(scripts): one implementation of per-architecture library selection

[datlechin]

Selecting the per-architecture slice of a vendored static library existed in five places that disagreed with each other. It is now one function in scripts/lib/macos.sh, the library #2336 introduced.

What was there

build-release.sh carried four near-identical functions, 159 lines between them:

	lines	how it resolved a library
prepare_mariadb	55	cd Libs, thin the universal only
prepare_libpq	30	thin the universal only
prepare_libmongoc	37	per-arch file, else thin the universal
prepare_hiredis	37	per-arch file, else thin the universal

and scripts/ci/prepare-libs.sh did a bare cp "Libs/${lib}_${ARCH}.a" with no fallback at all. So the release path and the CI step took different sources for the same library, and two of the four could not use a per-architecture file even when one existed.

build-release.sh drops from 693 lines to 529.

A regression I introduced and then caught by running it

The shared function first keeps a library that is already correct, which the originals expressed as:

lipo -info "Libs/${lib}.a" | grep -q "$target_arch"

That is a substring test, and a universal library contains arm64, so it matches. Faithful to build-release.sh, but not to ci/prepare-libs.sh, whose unconditional cp always produced a thin library. Porting the substring form would have left fat libraries in place and quietly changed the release to link those.

Running it is what showed this: after the first version, prepare-libs.sh arm64 reported libpq.a x86_64 arm64. The check is now an exact architecture match, so both callers produce a thin library, which is what the release path has always effectively had.

Verified

scripts/ci/prepare-libs.sh was run for real, both ways:

arm64    libmariadb.a arm64    libpq.a arm64    libhiredis.a arm64    libmongoc.a arm64
x86_64   libmariadb.a x86_64   libpq.a x86_64   libhiredis.a x86_64   libmongoc.a x86_64

Running it twice for the same architecture is a no-op, and a missing library fails naming it rather than leaving a half-prepared tree. Libs/ was restored afterwards and re-verified against the checksum baseline committed in git.

The extracted function was diffed against the original prepare_hiredis operation by operation: same three branches in the same order, the only textual difference being $LIBS_DIR in place of a relative Libs/, which also makes it independent of the caller's working directory.

build-release.sh itself is not run here, because it signs and notarizes.

https://claude.ai/code/session_013MEaba8K1HQcyDNeq5wEFk

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.