recipes: flet-libwebp 1.6.0 + Pillow WebP support

[ndonkoHenri]

Adds libwebp and rebuilds Pillow against it, so Image.open() and Image.save() handle WebP on device. Requested in flet#6806.

• Recipe

Our Pillow shipped PIL/_webp.pyi and no PIL/_webp.*.so, so opening a WebP warned WEBP support not installed and then raised UnidentifiedImageError. The requester's diagnosis was correct.

Why not the webp package that was asked for

The request names webp (anibali/pywebp). Building it would have satisfied the title and not the complaint: it registers no PIL plugin — zero register_open / register_extension calls in webp/__init__.py — so PIL.Image.open() would still have raised the same error. It also declares conan>=2.0 in build-system.requires, with an arch list containing no Android ABI and the target derived from the build host's platform.machine(), so cross-compiling it means patching conan out and injecting our own libwebp — i.e. writing flet-libwebp anyway and then bolting a second binding on top.

What declining it costs: pywebp's byte-budget encoding (WebPConfig.target_size / passes) and zero-copy decode into a preallocated numpy array. Animated WebP encode and decode were the strongest argument for it, and they don't survive reading the source — WebPImagePlugin.py binds WebPAnimDecoder/WebPAnimEncoder, so Pillow already has both.

Recipe shape

Autotools, not CMake, for an Android reason rather than a stylistic one: libwebp's CMakeLists calls set_version() unconditionally, producing libwebp.so.7, which APK packaging will not carry. configure goes through libtool, whose linux*android* case sets version_type=none and yields a bare libwebp.so. The Google release tarball is a make dist artifact with a pre-generated configure, so there is no autoreconf/autopoint step.

1.6.0 is not the newest for its own sake — it is the version Pillow 12.2.0 pins in depends/install_webp.sh, so we track what upstream tests against.

Two lines in build.sh are load-bearing, each having fixed a real failure:

• --prefix at configure time, not make install prefix=. The latter passed on all three iOS slices and failed on all three Android ones with cannot install `libwebp.la` to a directory not ending in /usr/local/lib — iOS is static-only so libtool never relinks, while Android's libwebp.la links libsharpyuv.la and trips libtool's rpath consistency check.
• hardcode_into_libs=no. Without it every Android .so carried RUNPATH /Users/<me>/…/wheel/opt/lib, which on CI would publish /home/runner/work/… inside the wheel. A grep guard follows the sed, because the silent failure mode is a published wheel with build-machine paths in it.

Android ships the shared libraries and iOS the static ones, matching how each resolves them. That split also disposes of sharpyuv without a consumer patch: on Android libwebp.so declares libsharpyuv.so as its own DT_NEEDED so symbols resolve transitively (llvm-readelf shows 0 sharpyuv references in _webp.so), and on iOS Pillow's setup.py already appends -lsharpyuv under sys.platform == "ios".

license_file names COPYING and PATENTS explicitly. Auto-discovery would find COPYING and silently drop PATENTS, which every libwebp source header points at as the additional patent grant.

Pillow

Two lines: the host dep and build.number: 2. No patch change. WebP was never force-disabled the way brotli/raqm/fribidi are in setup-12.x.patch; it was off only because feature detection found no libwebp, and the recipe already feeds CPATH/LIBRARY_PATH at {platlib}/opt/{include,lib} on Android while iOS resolves through sysconfig.

Nine tests, network-free, with three fixtures totalling 382 bytes. They prove more than linkage: lossless decode asserts exact quadrant colours, lossy decode allows a VP8 tolerance, and the round-trips check the RIFF container tag (VP8L lossless, VP8X for the alpha case). The EXIF test exercises libwebpmux and the animation tests exercise libwebpdemux and libwebpmux, so all three sub-libraries are proven at runtime.