Expose real BaSiCPy correction through CPU-only compiled workflows #320

Parent owns this as part of PR217. Current source-qualified MCP acceptance with actual fork wheel1.3.1 exposed two distinct boundaries: development-client children intentionally do not inherit a temporary PYTHONPATH wheel target, so the real acceptance environment must install the distribution; independently the CPU-only native registry admits only NumPy array contracts, while basic_flatfield_correction_jax declares JAX input/output storage. BaSiCPy public fit_transform accepts and returns NumPy; internal solver JAX does not require JAX array transport or a GPU. Source witness: OpenHCSRegistry _allowed_openhcs_memory_types, adapter @jax_func, ArrayBridge stack_slices GPU requirements. Do not substitute the NumPy approximate BaSiC operation or broaden GPU admission. Acceptance: genuine installed candidate dependency; CPU-only catalog exposes the exact real algorithm; public MCP compiled24-site same-channel fit persists float corrected images and same-fit flatfield/darkfield with correct source/calibration/contributor provenance. Retain failed receipts and verify owned-process closure. Original fixture/receipts: /home/ts/wt/openhcs-issue-batch-20260929/basicpy-live-wheel-20260930.