Three layers of test. All three are required for any change to accuracy- or latency-critical code.

1. Property tests (fast, per-PR)

Manifold identities are property-tested with proptest — see architecture-manifolds for the required identities. Run:

make test

2. Synthetic GT suite (slower, nightly)

Blender-rendered scenes with known 6DoF, point, contour, and face trajectories. Each tracker has a minimum accuracy bar that regressions must not cross:

• SE(3) pose: rotation error < 0.5°, translation error < 1% over 95% of frames.
• Point tracker: mean 2D error < 0.5 px.
• Face pose: rotation error < 2°.

Scenes live at tests/synthetic/scenes/. Ground truth is serialized alongside each render.

3. Criterion benches (latency)

make bench

Targets:

• uify-core::filters::ekf_manifold::bench_step — < 10 µs per step.
• uify-runtime::ringbuf::bench_read — < 100 ns per read.
• End-to-end uify-clap-plugin::bench_frame_cycle — budget per stage.

Live profiling

Tracy is wired in via tracy-client. To enable in a dev build:

make CARGO_FEATURES="tracy"

A Tracy session visualizes per-frame timing for every stage on every thread. Use this to diagnose latency regressions before the bench notices them.