Crate: uify-runtime. Owns all I/O and real-time bridging.

camera::

The Camera trait + Frame<'a> lend-out shape:

pub trait Camera {
    type Error;
    fn next_frame(&mut self) -> Result<Frame<'_>, Self::Error>;
}

pub struct Frame<'a> {
    pub t: Timestamp,
    pub width: u32,
    pub height: u32,
    pub format: FrameFormat,         // Rgba8 | Bgra8 | Gray8
    pub data: &'a [u8],
}

Reference impl: SyntheticCamera produces fixed-cadence frames with per-frame index bytes — useful for tests, demos, and offline pipelines.

Real backends are not yet implemented. Planned:

Backends will deliver zero-copy where the OS supports it (CVPixelBuffer on Apple, dma-buf on Linux). Cross-thread frame ownership will need a FrameHandle type beyond what Frame<'a> covers — to be added when a backend lands.

inference::

The Inference trait:

pub trait Inference {
    type Output;
    type Error;
    fn infer(&mut self, frame: &Frame<'_>) -> Result<Self::Output, Self::Error>;
}

Reference impl: ConstantInference<O> returns a fixed clone — useful for wiring downstream code against a deterministic upstream signal.

ONNX Runtime + per-platform EPs (CoreML / DirectML / NNAPI / XNNPACK / TensorRT / WebGPU) are not yet implemented. The Cargo.toml reserves [features] flags (onnx, coreml, directml, nnapi) for when they land.

ringbuf::

Wait-free SPSC ring of Sample<G, C> values. Single writer (vision thread), single reader (audio thread). Built on rtrb.

use uify_runtime::ringbuf::{channel, RingbufError};
use uify_core::Sink;

let (mut tx, mut rx) = channel::<Vector2<f64>, Matrix2<f64>>(64);

// Producer side: implements `Sink<G, C>`.
tx.write(&sample)?;                  // RingbufError::Full if consumer is behind

// Consumer side: wait-free pop.
while let Some(s) = rx.pop() {
    // process s
}

Allocation-free after channel is constructed for fixed-size geometries (Vector2, Vector3, Bbox, Se3). Variable-size geometries (contours, landmark sets) will need an arena strategy — not yet implemented.

RingbufError::Full policy: the new sample is dropped (drop-newest). Drop-oldest would mask consumer starvation; drop-newest at least surfaces it in logs. See architecture-threading for the full real-time discipline.