README.md

Tests: CPU, GPU simulation, and hardware

Directory	Runs where	What it checks
cpu/	CPU only	Core contracts, indicators, raw artifacts, CLI, example reference, mocked RunPod lifecycle, benchmark CPU baseline, import boundaries
gpu/test_simulator.py + simulator_cases.py	CPU CUDA simulator in a child process	Actual GPU kernels against explicit matrix/RSI answers and CPU reference
gpu/test_hardware.py	Real NVIDIA GPU	Kernel reductions, RSI reference and single-combination raw output; marked gpu

python -m pytest                         # CPU + isolated simulation; no rented pod
python -m pytest tests/cpu               # CPU tests only
python -m pytest tests/gpu/test_simulator.py
NUMBA_ENABLE_CUDASIM=0 python -m pytest -m gpu -v

test_simulator.py launches simulator_cases.py with CUDASIM in a separate process. The parent keeps its original environment; simulation cannot silently contaminate later hardware checks. The case file intentionally has no test_ prefix, so it is collected only by the child process. Hardware tests refuse simulation and skip if no real GPU exists; actual passes are required for hardware validation.

Hand-calculated trades anchor timing/fees independently. A known algebraic matrix anchors sums and float32 sums of squares. GPU kernels are compared with a Python reference; the benchmark's parallel CPU baseline is likewise checked independently. RunPod tests inject allocation/SSH/job/download/cleanup failure, cancellation, price rejection, ambiguous replies, metadata failure, and unsafe archives without leasing any resources. Architecture tests prevent examples/tools from entering core.

The optional installed hardware smoke tool is gpu-backtest gpu-check; its code is in tools/gpu_backtest_tools/checks/, not the engine. Performance measurements live in benchmark evidence. The original small hardware validation is retained under results.

For changes to numerical logic, run both simulation and real-GPU checks. A pure module/layout refactor must preserve the kernel body and compare previous/current outputs. Installing the wheel and checking its RunPod export catches packaging errors that editable-checkout tests can miss.

Raw output tests reload NPZ with pickle disabled, assert exact float64 array values, ordered dimensions, fees, schema-2 counts and file SHA-256. Single-combination output is anchored against hand-calculated profitable, fee-paying and losing trades. Tests also reject removed analysis commands/config fields before RunPod allocation.