[pumano]

Summary

Gets the example app building and its integration tests passing on the iOS
simulator using pure Swift Package Manager (no CocoaPods), and adds an
integration test suite that exercises every ML Kit plugin without a physical
camera.

On the iOS simulator the suite runs as +13 ~3 — 13 passed, 3 skipped,
0 failed.

The problem

feature/spm depended on mdata-group/google-mlkit-swiftpm at
9.0.0-simfix2. That package is unusable for a simulator build:

• arm64-only simulator slices — the xcframeworks had no x86_64 simulator
  slice, so the universal build the Flutter tool forces
  (-destination "generic/platform=iOS Simulator") failed.
• 33 KB stub binaries — the framework binaries were stubs, not the real ML
  Kit code.
• No models — recognizers crashed at runtime with "Invalid model path".

mdata-group/google-mlkit-swiftpm is itself a fork of
d-date/google-mlkit-swiftpm. Switching to the upstream d-date package at a
proper release fixes all three at once.

The solution: d-date/google-mlkit-swiftpm 9.0.2

d-date repackages upstream ML Kit 9.0.0 as SPM binary-target
xcframeworks. Since 9.0.1 it:

• ships proper universal slices — ios-arm64 (device) and
  ios-arm64_x86_64-simulator. The arm64 simulator slice is synthesised from the
  device slice by rewriting the Mach-O platform, so no EXCLUDED_ARCHS
  workaround is needed and the Flutter universal build works;
• includes the models inside the xcframeworks for text recognition, pose
  detection and selfie segmentation;
• is App Store Connect compatible.

Because SwiftPM cannot ship resource bundles inside a binary target, the model
resource bundles must be added to the app manually (see below).

What changed

Package manifests (11 Package.swift) — switched from
mdata-group … 9.0.0-simfix2 to d-date/google-mlkit-swiftpm exact: 9.0.2:
barcode_scanning, commons, digital_ink_recognition, entity_extraction,
face_detection, image_labeling, language_id, object_detection, pose_detection,
selfie_segmentation, smart_reply, text_recognition, translation.

Example app (packages/example)

• ios/Runner.xcodeproj/project.pbxproj — adds the 19 ML Kit resource bundles
  (file refs + an "MLKit Bundles" group + Resources phase) and -all_load in
  the linker flags.
• ios/mlkit_bundles/ — the 19 bundles, gitignored (≈50 MB of model files
  are not committed).
• ios/download_mlkit_bundles.sh — new; fetches the bundles from the d-date
  release.
• ios/.gitignore — ignores mlkit_bundles/.
• ios/…/swiftpm/Package.resolved (×2) — pin d-date 9.0.2.
• ios/Runner/AppDelegate.swift — registers a tiny example/simulator method
  channel so tests can detect the simulator.
• pubspec.yaml / pubspec.lock — integration_test dev-dependency.
• README.md — documents the one-time bundle download step.
• integration_test/ml_kit_integration_test.dart — new; the test suite.

Plugin fix

• google_mlkit_entity_extraction was a silent no-op: it only depended on
  commons, so #if canImport(MLKitEntityExtraction) was false and it returned
  UNIMPLEMENTED. Added the MLKitEntityExtraction product so it genuinely
  works.

How it works (technical notes)

• The real ML Kit code lands in the app binary, not the frameworks. With
  -ObjC + -all_load, every object file in the ML Kit static archives is
  pulled into Runner.debug.dylib (≈530 MB in debug). The per-framework
  binaries copied into Runner.app/Frameworks/ are small stubs; the symbols
  (MLKTextRecognizer, MLKBarcodeScanner, MLKFaceDetector, …) live in the
  app dylib. This is expected for statically-linked ML Kit.
• -all_load is required. Without it the ML Kit initializers (which
  register models/resources via +load) are dead-stripped and recognizers fail
  at runtime.
• Plugin consistency. Every plugin that has an available d-date product is
  wired to it. The remaining stubs are correct because d-date 9.0.2 ships no
  such product: document_scanner, face_mesh_detection,
  subject_segmentation, and the six genai_* plugins.

Integration tests

integration_test/ml_kit_integration_test.dart covers every plugin. Two input
helpers avoid needing a real camera:

• inputImageFromAsset(path) — copies a bundled asset to a temp file and builds
  InputImage.fromFilePath (the file-path native path).
• bitmapFromAsset(path) — decodes an asset to raw RGBA and builds
  InputImage.fromBitmap (the same shape a camera frame produces).

Group	Input	Simulator behaviour
barcode scanning	asset file	runs
text recognition	asset file	runs
face detection	asset file	runs
face mesh detection	asset file	asserts MissingPluginException (no iOS API)
image labeling	asset file	runs
object detection	asset file	runs
pose detection	asset file	skips (MediaPipe GPU)
subject segmentation	asset file	asserts MissingPluginException (no iOS API)
selfie segmentation	asset file	skips (MediaPipe GPU)
digital ink recognition	synthetic stroke	runs
language identification	string	runs
entity extraction	string	runs
translation	string (downloads model)	runs
smart reply	conversation	runs
camera frame pipeline (emulated)	asset → bitmap	runs
camera (smoke)	real camera	skips (no camera / no permission)

All async ML Kit calls run inside tester.runAsync(...).

Why 3 tests are skipped

The 3 skips are all simulator limitations, not code problems:

Skipped test	Why it skips on the simulator
pose detection	ML Kit pose detection is MediaPipe-backed. The simulator has no real GPU, so MediaPipe's GL path hits LOG(FATAL) in mediapipe::GpuBufferStorageCvPixelBuffer::GetTexture → abort(). That is a hard crash of the whole app that cannot be caught from Dart, so the test detects the simulator (via the example/simulator channel) and skips. Runs on a real device.
selfie segmentation	Same reason — also MediaPipe-backed (same GL path) → same crash → skipped. Runs on a real device.
camera (smoke)	The simulator exposes no camera: availableCameras() is empty / initialize() throws (no permission granted). The test skips.

Note that face mesh and subject segmentation are not skipped: ML Kit has
no iOS API for them, so the plugins are graceful stubs returning
FlutterMethodNotImplemented. Those two tests pass on iOS by asserting the
MissingPluginException (and run the real detection on Android).

How to test

cd packages/example

# one-time: fetch the (gitignored) ML Kit model bundles
cd ios && ./download_mlkit_bundles.sh && cd ..

# ensure the Flutter SPM flag is on (one-time)
flutter config --enable-swift-package-manager

flutter build ios --simulator

flutter test integration_test/ml_kit_integration_test.dart -d <SIMULATOR_UDID>

flutter analyze on the example app: No issues found.

Caveats

• Debug binary is large (≈530 MB Runner.debug.dylib) because -all_load
  pulls in all of ML Kit. Fine for tests; a release build is smaller.
• Runtime model downloads. Digital-ink, entity-extraction and translation
  download their models on first use, so those tests need network access.
• MediaPipe-backed modules (pose, selfie segmentation) are untestable on the
  simulator and must be verified on a real device.

[pumano]

@fbernaly Could you check it via real iphone?

[fbernaly]

@pumano : I tested on a physical device and everything looks good! I just pushed a few dependency version updates.

We're ready to merge this into develop and release a new plugin version. You've put a lot of work into this, so please branch off the latest code from feature/spm (including my latest commit) and open a PR against develop—I'd like you to get the credit for your contributions.