Performance-Gated Tests

SC-NeuroCore keeps timing sanity checks out of the default pytest path. Tests that assert a wall-clock threshold are marked with SC_NEUROCORE_PERF=1 so regular local and pull-request runs stay deterministic and cheap. The Performance Benchmarks workflow runs only by manual dispatch with a required run reason and change reference. It runs the same perf-gated selector alongside the Rust and Python benchmark suites. Run it for a new catalogue model after fidelity gates, a material numerical or compute-path change, a confirmed regression, a planned stable baseline, or an explicitly scoped comparison. Documentation changes, an unrelated source hash, and routine CI repetition do not justify a new timing record.

The Rust Criterion comparison still posts an alert for a large timing change, but that alert is advisory because runner hardware and host state affect the measurement. Build, execution, and artifact-upload errors still fail the workflow. The dispatch reference and source SHA are retained with the raw benchmark output; historical timing records are not overwritten to match new code. The planned shared benchmark utility will add sealed plans, before/after hardware-state capture, immutable result lineage, and the same trigger decision for Catalogue and Studio before SC-NEUROCORE STABLE.

Run the same selector locally when changing a perf-gated file:

Bash

SC_NEUROCORE_PERF=1 PYTHONPATH=src:. python -m pytest \
  tests/test_export/test_onnx_exporter_performance.py \
  tests/test_learning/test_lifelong.py \
  tests/test_learning/test_federated.py \
  tests/test_transformers/test_block.py \
  tests/test_layers/test_recurrent.py \
  tests/test_layers/test_vectorized_layer_packed_forward.py \
  tests/layers/test_sc_conv_layer_configuration.py \
  tests/test_layers/test_sc_dense_layer_performance.py \
  tests/test_layers/test_memristive.py \
  tests/test_layers/test_sc_learning_layer.py \
  tests/layers/test_fusion_performance.py \
  tests/test_hdc/test_base.py \
  tests/accel/test_vector_ops_validation_performance.py \
  tests/test_solvers/test_ising.py \
  tests/test_optics/test_photonic_layer.py \
  tests/test_quantum/test_hybrid.py \
  tests/interfaces/test_dvs_input_performance.py \
  tests/sources/test_bitstream_current_source_performance.py \
  tests/hdl_gen/test_verilog_generator_io_perf.py \
  tests/test_hdl_gen/test_spice_generator.py \
  tests/test_bio/test_dna_storage.py \
  tests/test_bio/test_grn.py \
  tests/test_graphs/test_gnn.py

Current Selector

Surface	Test file
ONNX export	tests/test_export/test_onnx_exporter_performance.py
Lifelong learning	tests/test_learning/test_lifelong.py
Federated learning	tests/test_learning/test_federated.py
Transformer block	tests/test_transformers/test_block.py
Recurrent layer	tests/test_layers/test_recurrent.py
Vectorized layer	tests/test_layers/test_vectorized_layer_packed_forward.py
Convolution layer	tests/layers/test_sc_conv_layer_configuration.py
Dense layer	tests/test_layers/test_sc_dense_layer_performance.py
Memristive layer	tests/test_layers/test_memristive.py
Learning layer	tests/test_layers/test_sc_learning_layer.py
Fusion layer	tests/layers/test_fusion_performance.py
HDC encoder	tests/test_hdc/test_base.py
Vector operations	tests/accel/test_vector_ops_validation_performance.py
Ising solver	tests/test_solvers/test_ising.py
Photonic layer	tests/test_optics/test_photonic_layer.py
Quantum hybrid layer	tests/test_quantum/test_hybrid.py
DVS input layer	tests/interfaces/test_dvs_input_performance.py
Bitstream current source	tests/sources/test_bitstream_current_source_performance.py
Verilog generator	tests/hdl_gen/test_verilog_generator_io_perf.py
SPICE generator	tests/test_hdl_gen/test_spice_generator.py
DNA storage	tests/test_bio/test_dna_storage.py
Gene-regulatory network	tests/test_bio/test_grn.py
Graph neural layer	tests/test_graphs/test_gnn.py

The selector is intentionally pytest-based. It is regression evidence for small timing thresholds inside existing behaviour tests, not a replacement for isolated benchmark artefacts under benchmarks/results/. Public benchmark claims still require the benchmark-evidence rules described in Benchmarks.