Skip to content

recipes: version bumps, part 1 - #135

Merged
ndonkoHenri merged 12 commits into
flet-dev:mainfrom
ndonkoHenri:recipe-bumps
Nov 14, 2026
Merged

ndonkoHenri merged 12 commits into
flet-dev:mainfrom
ndonkoHenri:recipe-bumps

Conversation

@ndonkoHenri

Copy link
Copy Markdown

What

Part 1 of bringing recipes up to their latest upstream releases: ten recipes, one commit each.

Recipe Bump Notes
msgpack 1.1.2 → 1.2.4 Security fixes GHSA-j586-36cw-2gc2, GHSA-j8c6-9vc5-hj6x. New test that the C extension is in use, since msgpack silently falls back to pure Python.
pydantic-core 2.47.0 → 2.50.4 pydantic 2.14.x pins exactly 2.50.x, so it can't install today.
aiohttp 3.14.0 → 3.14.5
markupsafe 3.0.3 → 3.0.4
xxhash 3.8.0 → 4.1.0
kiwisolver 1.5.0 → 1.5.1
pyzmq 27.1.0 → 27.3.0
pillow 12.3.0 → 13.0.0 setup-13.x.patch: 13.0 reworked the sys.prefix lookup our patch edits. 12.x keeps its own.
ncnn 1.0.20260526 → 1.0.20261020 New skip-upstream-ios-cross.patch, see below.
curl-cffi 0.16.2 → 0.16.3 With flet-libcurl-impersonate 2.1.1 → 2.2.2. Not 0.17.0, see below.

Why not the latest everywhere

curl-cffi stops at 0.16.3. 0.17.0 needs curl-impersonate 2.3.x, whose iOS build resolves DNS through c-ares. On the iOS simulator every request then fails with curl: (28) Resolving timed out. The same app on 0.16.2 makes 5 of 5 requests. Upstream moved Android back to the threaded resolver (lexiforest/curl-impersonate#339, #344) but not iOS. 2.2.2 still uses the threaded resolver, and the example app makes 5 of 5 requests on 0.16.3. flet-libcurl-impersonate now ships the licence notices from upstream's tarballs instead of vendored copies.

ncnn iOS on 3.13/3.14 is a known failure. 20261020 added a cibuildwheel iOS block to setup.py that swaps forge's compiler wrappers for raw xcrun clang++, so the simulator slice linked as device iOS. The patch skips that block when EXTRA_CMAKE_ARGS is set, and all our iOS slices now build as MH_DYLIB. But on cp313/cp314, PyPI's own ncnn wheels are tagged ios_14_0 against our ios_13_0, and the platform tag outranks our build tag. Their extension is MH_BUNDLE, which Flet's iOS link rejects. Any Flet that accepts ios_14_0 wheels already gets those upstream wheels, whatever version we ship. Android and cp312 iOS get ours. Follow-up: tag our ncnn iOS wheels ios_14_0 (the binary's minos is already 14.0).

Validation

On-device recipe tests on Android and iOS, Python 3.12–3.14:

  • msgpack: run 43604693413, 6/6 legs, 3 tests each
  • curl-cffi 0.16.3: run 43630791138, 6/6 legs, 4 tests each
  • ncnn + pydantic-core: run 43617825651. pydantic-core 6/6. ncnn Android 3/3, iOS 3.12 pass; iOS 3.13/3.14 fail as described above.
  • the other seven: run 43607221256, 6/6 legs each

curl-cffi's tests make no network calls, so its example app was also run on an iOS simulator against the CI wheels, making real HTTPS requests.

Two Android 3.12 jobs failed on a Gradle download in serious_python (a zero-filled ETag JSON) and passed on retry. That flake is unrelated to these recipes.

Bump 1.1.2 -> 1.2.4 (security fixes GHSA-j586-36cw-2gc2, GHSA-j8c6-9vc5-hj6x).
Add a test that the C extension, not msgpack's silent pure-Python
fallback, is the implementation in use.
20261020's cibuildwheel iOS block swaps in raw xcrun clang++, which links the
simulator slice as device iOS. The recipe already passes those settings.
curl-impersonate 2.2.3+ resolves DNS through c-ares on iOS, and every lookup
times out on the iOS simulator (curl 28); 0.16.2 makes the same requests in the
same app. Android is unaffected upstream, but one version for both keeps the
recipe simple. armeabi-v7a stays excluded: 2.2.2 has no 32-bit ARM asset.
The tarballs ship their own licence notices, so the vendored copies go.
@ndonkoHenri
ndonkoHenri merged commit 9ee61c5 into flet-dev:main Nov 14, 2026
2 of 4 checks passed
@ndonkoHenri
ndonkoHenri deleted the recipe-bumps branch November 14, 2026 19:46
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant