Baseline lock and regression net

Captured on 2026-06-14 with the unmodified master tree.

• Test target: ctest -R golden_baseline --output-on-failure
• Full-suite target: docker compose run --rm dev ctest --output-on-failure
• Benchmark snapshot will be recorded here once a golden benchmark manifest exists.

Captured test list (U0)

CTest Name	Status
sanity.*	passed
elm_cpu.*	passed
os_elm.*	passed
os_celm.*	passed
h_os_elm.*	passed
rbf_features.*	passed
demo_api.*	passed
cuda_infra.*	skipped (CPU-only)
elm_gpu.*	skipped (CPU-only)

Golden baseline expectations

• BatchElm with fixed weights/biases and deterministic input targets reproduces a bounded prediction MSE on training data.
• OsElm after initialization reproduces bounded prediction MSE on the same data.
• Online OsElm::update improves fit on held-out data within the same bounds.
• The golden test file is the regression anchor for U1 onward. Any later step must keep these three cases green.

Regression net

Use this document to record when a refactor changes numeric behavior. Updates should include the step that triggered the change, a brief justification, and a new threshold if the old one is no longer appropriate.