rk_infer 0.2.1

On-device inference for Dart over an embedded C++ engine, with the raw frame never leaving the worker that owns it.

0.2.1

• macOS and iOS actually build now: the pod script phase shipped with CRLF line endings and exited 0 without invoking cargo. rk_infer_task_names is also declared in the C header, which it was exported without.

0.2.0

• The package is now a Flutter FFI plugin: flutter build runs cargo itself and puts the library into the application on Windows, Linux and Android. Before this there was no flutter.plugin.platforms block in the pubspec, and the native part reached no build at all.
• Fixed along the way: the cfg(unix) branch of rust/src/dylib.rs compiled on neither Linux nor Android, and the very first build for a target other than Windows caught it. macOS and iOS are written but have never been built.

0.1.0

The contract, the native side, and the model delivery chain. What a consumer gets and what it does not:

• InferenceEngine, backed by NativeInferenceEngine.start(...), which runs the engine in a long-lived worker isolate. Every method returns a Future; there is no synchronous entry point, so inference cannot land on the isolate that runs the interface.
• run takes a TransferableTypedData — a move, not a copy. Nothing on the contract returns frame bytes, writes them to disk, or opens a socket.
• Failures are values: a sealed InferError with a case per native status, mapped by name, and UnknownNativeStatus for a name this build does not know. Nothing throws, including a panic inside the native library.
• Models load from a manifest delivered by apt, and are refused unless the schema, the ABI, the retention, the declared length, the SHA-256 and the version pin all agree. See doc/models.md.
• InferenceOutput carries its own retention deadline. A manifest without a finite, non-zero retention does not load.
• Two tasks, deliberately: visitorCount (a fallback, where the camera does not already count) and unscannedItemHint (open research — a filter that narrows what a person reviews, not a detector of theft).

Not in this release: running a model. loadModel verifies everything and then reaches NotImplemented, whose message names the ONNX Runtime entry points still to be bound. There is deliberately no stub returning an empty result, because on a quiet camera that is indistinguishable from a working engine.

No figure in this package has been measured on the hardware it is for.

Changed since 0.0.1:

• hasNativeEngine now reports true. It was documented as "always false for now" and answers a question about the package, not the machine — it does no native work, because a synchronous probe would be the very call on the interface isolate this design forbids. For the machine's answer, start an engine and read EngineCapabilities.

0.0.1

• Name claimed. No content yet: the package proves the publishing pipeline, it does not solve the problem.