Tests and validation

Use the lane scripts: they select the right backends, fixtures and build directories. Do not run a raw cargo test or pytest to validate a change in the repository. What each lane runs, what it costs and its pitfalls are in the test lanes in detail.

Choosing a lane

Change	Command to run
Everyday Rust	scripts/test-fast-rust.sh
Python API	scripts/test-fast-python.sh
HTTP handler	scripts/test-server.sh while iterating; test-fast-rust.sh before shipping
Runtime, ABI or FFI	scripts/test-abi.sh
Training or CPU integration	scripts/test-cpu-integration.sh
Recurrent or hybrid architecture	scripts/test-cpu-integration.sh, with a model per family
Container integration	scripts/test-container.sh
RIR kernel, emitter or schedule	scripts/test-rir-parity.sh
Promoting an RIR kernel	scripts/test-rir.sh
Removing a native RIR kernel	scripts/test-rir-graph.sh

tests/recurrent_families.rs runs inside the CPU integration lane and covers one family of recurrent state per row, not one model: shortconv is the default fixture and always runs, conv_ssm and gated_delta_net skip unless RETRO_FALCON_H1_TEST_MODEL and RETRO_QWEN3NEXT_TEST_MODEL point at a GGUF. Touching the packed-training capability, the micro-batch finiteness check or the recurrent rollback derivation means running that lane with all three set - a lane that only ever loads the default fixture proves nothing about the other two families.

test-fast-rust.sh and test-fast-python.sh are the minimal checks for a change that touches both layers. Lanes that need a GPU, a model or a daemon say so and must be run before a PR when their scope is affected.

RIR: iterate, then decide

To adjust a shader or a geometry, start with:

scripts/test-rir-parity.sh

This lane checks the generated shader against the Loop IR oracle, without a model. To quickly measure a given schedule, use the RIR timing test:

RIR_TIME=1 cargo test --release -p rir-runtime --test device_timing -- --nocapture

An offline search can propose geometries:

cargo run --release -p rir-sweep -- --kernel <kernel-name>

Neither of those two commands promotes anything. Only scripts/test-rir.sh compares the generated kernel and the native kernel, on the same shapes and the same hardware; it is therefore what decides a registry policy (see RIR kernel promotion).

Build trees

The build directories are deliberately separated by backend and profile: sharing them can trigger needless llama.cpp recompilations. Each script's variables let you override them if needed.

Before a PR

Run the fast lane matching your layer and any lane listed in the table for the components you changed. Add the GPU lane for a backend change; add the ABI lane for the runtime or the FFI. A lane failure is the starting point for diagnosis: keep its log rather than retrying with a different raw command.