[boydjohnson]

Builds mapnik from source with the osmflat datasource compiled in as a built-in plugin, and links render against it. The result loads nothing at runtime — no libmapnik, no osmflat.input, no system fonts.

cmake -S . -B build-static -DCMAKE_BUILD_TYPE=Release \
    -DOSMFLAT_STATIC_RENDER=ON -DOSMFLAT_FULLY_STATIC=ON   # -FULLY_ is musl-only
cmake --build build-static --target render

The dynamic build is untouched and remains the default.

platform	runtime deps	archive
x86_64-linux-musl	none (fully static)	10.9 MB
aarch64-linux-musl	none (fully static)	10.2 MB
aarch64-apple-darwin	/usr/lib + /System only	22 MB

All three render byte-identical output to the dynamic build (md5 d48d0556925e17e872d0ac25229319d2 on the style-relations fixture).

How osmflat gets into a static libmapnik

Mapnik's static-plugin table is compile-time (src/datasource_cache_static.cpp + src/CMakeLists.txt), so cmake/static-mapnik/patch-mapnik.cmake edits the fetched tree: it adds plugins/input/osmflat (sources referenced from this repo, no copies), links it into libmapnik, and registers it in the static datasource map. Each edit anchors on an exact upstream string, fails loudly if mapnik moved it, and is skipped when already applied.

The scale_denominator trap

Dropping PROJ keeps a 9 MB proj.db out of the binary, but a PROJ-less mapnik knows only epsg:4326 and classifies it as geographic — making scale_denominator ~111319x larger than what PROJ reports for the equivalent +proj=longlat +datum=WGS84 +no_defs. Every style here is tuned against the degree-based numbers (0.1 ≈ neighborhood), so that would have silently disabled every MaxScaleDenominator gate while still rendering something plausible. (First attempt did exactly this: 214688 where the dynamic build says 1.92857.)

The patch instead makes a PROJ-less projection classify +proj=longlat exactly as PROJ does. Other proj4 strings still throw.

Platform wrinkles

musl: harfbuzz's CMake config names the shared library (attempted static link of dynamic object), so the pkg-config fallback is forced; cargo asks for -lgcc_s, which musl ships only shared, so Corrosion's link interface gets gcc_eh.

macOS: Homebrew's libharfbuzz.a is built against graphite2, which has no static archive — so harfbuzz is built from source with just freetype and CoreText (dropping glib/pcre2/intl too). icu4c/zlib/bzip2 are keg-only, so their prefixes are passed explicitly. Homebrew's Boost::regex names icu libs bare, which wouldn't resolve and would prefer dylibs; mapnik's own workaround drops them. The build script fails if otool -L shows anything outside /usr/lib or /System.

CI

scripts/build-static-render.sh (Alpine, via docker run — identical to a local podman run) and scripts/build-static-render-macos.sh build, test and package; scripts/package-render.sh is shared. static-render.yml runs all three platforms on every PR and caches the build tree — mapnik is ~16 min cold on x86_64, ~1 min warm.

release.yml builds the same three archives from a clean tree (no cache) on a v* tag, requires the tag to match project(... VERSION), smoke-tests each unpacked linux archive on a bare alpine image, and publishes with a SHA256SUMS. workflow_dispatch builds without publishing.

Licensing

The repo had no license file; it's now MIT. Apache-2.0 was considered and dropped because it's incompatible with LGPL-2.1 for a combined work, and the static binary links LGPL mapnik. Each archive ships LICENSE-MIT, mapnik's LICENSE.mapnik, and a NOTICE naming every statically linked library plus how to rebuild and relink.

Not addressed: x86_64 macOS (the matrix could gain it; only arm64 is built), and a warm-cache CI measurement — the key naming changed in the last commit, so the first restore happens on the next push.

🤖 Generated with Claude Code