Every transport implements uify_core::Sink<G, C> for the Gs it makes sense for. The trait is fn write(&mut self, sample: &Sample<G, C>) -> Result<(), Self::Error>.

OSC — uify-transport-osc

Implemented: OscPoint2DSink for Sample<Vector2<f64>, Matrix2<f64>>. Not yet implemented: bbox, pose, landmark, contour sinks.

OscPoint2DSink

use uify_transport_osc::OscPoint2DSink;
use uify_core::Sink;

let mut sink = OscPoint2DSink::new(
    "127.0.0.1:0",                   // local: OS picks port
    "127.0.0.1:9000",                // remote: OSC listener
    "/uify/point/2d",                // OSC address
)?;
sink.write(&sample)?;                // any Sample<Vector2, Matrix2>

Each call to write sends one OSC message over UDP:

/<path>   ,hfff   <t_ns>   <x>   <y>   <confidence>

The timestamp payload is the host-monotonic nanosecond value, not an OSC time-tag. The same monotonic clock must be used at all ends of the pipeline.

Vector2<f64> components are cast to f32 on the wire — losing about seven decimal digits, fine for image-plane pixel coordinates. If a future geometry needs full f64 precision, it'd be a separate sink type using OscType::Double.

The covariance field on Sample is not sent. It's available for in-process consumers via the Sink trait but isn't part of the OSC schema.

MIDI / MPE — uify-transport-midi

Stub. The crate exists in the workspace but lib.rs is a placeholder. Planned: per-note expressive control via MPE, with hand-to-note mapping using per-note pitch-bend (X), timbre (Y), and pressure (Z); note-on / note-off driven by a gesture FSM in the consumer.

Shared memory — uify-transport-shm

Stub. Planned: POSIX / Win32 shared-memory segment with a header + ring of Sample records, for same-machine consumers that need µs-scale latency across process boundaries. The intra-process equivalent is already implemented at uify_runtime::ringbuf::channel.