[ndonkoHenri]

What

cffi 2.1.1 with ffi.callback() working on iOS, plus two CI fixes: testing rebuilt flet-lib* wheels, and pinned Ubuntu runners.

Why

On iOS every ffi.callback() raised MemoryError: Cannot allocate write+execute memory for ffi.callback(), and variadic ABI-mode calls returned garbage on arm64. cffi uses libffi's closure API on Apple only when FFI_AVAILABLE_APPLE is defined, which only Apple's own macOS ffi.h does. With the upstream libffi our iOS builds use, cffi falls back to its own allocator, which maps writable+executable memory, and iOS refuses. libffi itself needs no such memory there (arm64 uses trampoline tables, which is why CPython's ctypes callbacks work).

Upstream fixed this in python-cffi/cffi#265, merged but not in any release yet.

Changes

• recipes/cffi: 2.1.1, build 1, new ios-ffi-closure-alloc.patch enabling ffi_closure_alloc / ffi_prep_closure_loc / ffi_prep_cif_var under TARGET_OS_IPHONE (Android's preprocessed source is byte-identical). mobile.patch regenerated for 2.1.1, no behaviour change.
  • 2.1.1 rather than rebuilding 2.0.0: PyPI ships official cffi 2.1.1 iOS wheels (cp313-cp315) without the fix, and they outrank 2.0.0. At the same version, our build tag wins.
  • The patch is a TARGET_OS_IPHONE variant rather than #265 verbatim: forge's non-interactive patch reverse-applies a hunk it sees as already present, so a future bump to a release containing #265 would silently remove the fix. This variant lands as dead code instead. Drop it once a release includes #265.
  • Tests: a C-invoked callback (qsort), a callback from a native thread, 3000 closures allocated and freed, and a variadic snprintf. Four of five fail on the old wheels.
• recipes/soundfile: file-object reads (sf.read(io.BytesIO(...))) now work on iOS. Removed the test asserting the iOS MemoryError, dropped the iOS skip, updated README, example README and a comment.
• CI: dist-test/ now bumps py3-none (build.sh flet-lib*) wheels to build 9999 too, and Android's loop also matches abi3. Before, a rebuilt flet-lib* at an already-published version and build number could lose to its published copy in the device test.
• CI: Ubuntu runners pinned to ubuntu-24.04 (what ubuntu-latest resolves to today). ubuntu-latest will move to 26.04, which would change the Android lane (NDK builds, KVM emulator) for every recipe at once; macOS is already pinned to macos-26.

Validation

• On-device CI, Android + iOS, Python 3.12/3.13/3.14 (packages=cffi:,soundfile:, prebuild_recipes=cffi): all green on the fork. Note soundfile's iOS tests need the new cffi, so they only pass when cffi is built in the same run or published.
• iOS simulator, before/after: callbacks raise MemoryError and snprintf returns garbage on forge 2.0.0-1 and PyPI 2.1.1; all pass with this build.
• Not run: physical iOS devices, x86_64 simulator.

Consumer notes

Apps get working ffi.callback() on iOS automatically once cffi 2.1.1-1 is on pypi.flet.dev; for soundfile that means in-memory reads work on both platforms. Details: recipes/soundfile/README.md.