[monkytheluffy]

Standalone one-file binaries (PyInstaller) for releases, plus the CI workflow to build them on every release publication.

What it does

.github/workflows/binary.yml (triggers: release: types: [published] + workflow_dispatch) builds the ppgrid console script as a self-contained one-file binary per platform:

• ubuntu-latest (linux x86_64), macos-latest (macos arm64), macos-13 (macos x86_64)
• Per leg: platform prebuilt of the C interp using ppgrid._cinterp's own cc recipe (the .so name carries the cinterp.c src hash, so a stale build is never loaded), PyInstaller with --collect-data pyproj + --collect-submodules pyproj (PROJ database), --collect-all rasterio (wheel-bundled GDAL, no PyInstaller hook), --hidden-import pandas (loaded via importlib.import_module, invisible to static analysis), --copy-metadata ppgrid (__version__ via importlib.metadata), --add-data for cinterp.c + the src-hash .so (the loader needs cinterp.c present to even look for the .so).
• Gates: --version + clean-env e2e (fresh HOME, no venv) on all legs; no-compiler e2e (PATH without cc/gcc/clang, so the bundled tier-1 .so is the only way the C interp can load) on ubuntu; byte-exact anchor gate vs the committed examples/melb (10m/500m, exact bench tier2 config: --value-col price --cap-km 10 --skip-calibration --workers 4) on ubuntu only — the anchors are Linux-anchored, macOS legs get the functional e2e.
• xz the binary as ppgrid-<ver>-<os>-<arch>.xz; upload to v<ver> via softprops/action-gh-release@v2 (action-release@v2 was renamed upstream; the successor keeps tag-based upload, overwrite by default). On workflow_dispatch with no matching release the build falls back to an actions/upload-artifact so it stays verifiable.

README: one new "Standalone binaries" paragraph (what it is, release-assets location, no-Python/no-GDAL claim, C interp fallback to bit-identical np.interp).

Goal-1 proof (built + tested on hydrogen, v0.5.0, PyInstaller 6.22.3, python 3.13.12)

• dist/ppgrid --version -> ppgrid 0.5.0, exit 0; --help exit 0
• clean-env e2e (env -i, fresh HOME, PATH without cc/gcc/clang, data/melb_houses.csv --res 500 --cap-km auto): exit 0, rasters written, no ~/.cache created (tier-1 bundled .so loaded, no JIT, no compiler touched)
• anchor gate, binary output vs committed examples/melb — all four byte-exact:
  • 10m/value.tif 74d76b9b3a01dbc2ce76bf11c8fcd67d2c8b8b08ad1720b17983c28e00a5db43
  • 10m/support_km.tif 3d0822e1f15b545e359e51f748686d0500053f09b80e646ef7bf22b636a6c47d
  • 500m/value.tif 4a30bf476a85c9d30bb46aa84aa3aa140c532af0d03fda1dc90a72bfda3d45a7
  • 500m/support_km.tif 530cc25f3d4c5ebe2ab17359baeac56c635bc2f0096a14d70d7e790efaad9346
• linux binary: ppgrid-0.5.0-linux-x86_64.xz, 179,993,860 bytes, sha256 3840181c350516d8066e705e69c0edfa68e4ecdf23b191fbe82acdba392976a
• Attached to the v0.5.0 release and at https://drop.junkyard.sh/ppgrid-0.5.0-linux-x86_64.xz

Recipe note: the committed-era prebuilt _cinterp_c-314334463687...so in the working tree is a stale JIT artifact (gitignored) whose src hash no longer matches cinterp.c (current hash 450abce519f7, after the darwin-FMA fix 976ff46); this build recompiles with _cinterp._build's exact cc line (-O3 -ffp-contract=off -fPIC -shared -fvisibility=hidden -I<include>), verified by tests/test_cinterp.py (37 passed).

macOS binaries will be produced by running this workflow on the v0.5.0 release after merge (re-publish the release, or workflow_dispatch from the tag).