Unadapted FLM physical food reference: all 24 declared trials

With NumPy installed, audit the extracted directory:
python audit_food_core_physical.py records --bundle core --output fresh-audit.json

Set OMP_NUM_THREADS, OPENBLAS_NUM_THREADS and MKL_NUM_THREADS to 1 before launch.
The audit replays saved neural states, geometry and commands; it does not rerun physics.
No training corpus, PyTorch or FlyGym installation is needed for this audit.
core/ contains the original bound manifest and four inference payloads. Its earlier
open-loop fixture is supplied in the separate food-core-interface.zip release, not here.
sources/ preserves all bound local sources, not a complete training repository.
Read sources/docs/FOOD-CORE-PHYSICAL-REFERENCE.md for the prior declaration.
Read docs/PHYSICAL-STATE-SEMANTICS.md before reusing body poses: c_head aliases
rh_tarsus5, and cached sensor poses differ from kinematics of saved qpos.
Keep the MIT code license and included graph/body provenance and component licenses.
There are zero food-adaptation updates; do not treat contacts as proof of learning.
