# SPDX-License-Identifier: Apache-2.0
"""CPU reference of TransFusion-L (Autoware lidar_transfusion) -- the ground truth every PCC / output-agreement gate
compares against (PLAN.md 0.3 item 1). It runs where ttnn is absent (the research venv) and never imports ttnn.

- ``config.py``    the Autoware parameters this port reproduces (the ml_package / remapper / metadata yaml of the
                   weights, ``transfusion.param.yaml``, the node's hard-coded constants and silent sanitization) and
                   the port's COMPILE constants (``N_MAX`` 65,536, sampling modes), each with its source line in
                   autoware_universe @ 9ceaccf026c31ffc5319bc9eeb4bd7bede0af3fd;
- ``weights.py``   ``transfusion.onnx`` read as data with the vendored ``ttaw.weights.OnnxWeights`` (initializers
                   and constants by consuming node; every wiring assumption asserted from the graph; BN folded in
                   float64 where the TT port folds it: the PFN MatMuls and the deblock2 ConvTranspose);
- ``model.py``     fp32 torch rebuild of the deployed graph (PFN, scatter, SECOND, SECONDFPN, dense head, local max,
                   ONNX-semantics top-K with optional teacher forcing, query init, decoder, prediction heads,
                   output assembly) with per-module taps (``ttaw.golden.TapRegistry``; names in ``MODULE_TAPS``);
- ``rewrites.py``  the TT-friendly equivalent forms (PFN tile form and 32-slot pillars, K-split shared conv,
                   deblock rewrites, C22 masked local max, arithmetic ``query_pos``, class table, constant cross-
                   attention K / V tables, head-dim padding, merged prediction heads), proven against ``model.py``;
- ``ort.py``       ONNX Runtime golden of the shipped ONNX on the same voxel tensors, with optional intermediates
                   (research / tests only: onnxruntime is not in the image);
- ``pipeline.py``  the reference end to end: ``host/`` pre-processing -> torch or ORT network -> ``host/``
                   post-processing -> ``ttaw.outputs.Detections3D``.

Pre- and post-processing are not duplicated here: they live in ``tt_transfusion.host`` (shared with the device path,
so TT and CPU see identical voxel and feature tensors) and are tested against the research scripts.

The ``to_dict()`` of this reference's output on the shipped sample is stored as
``samples/test_pcd.reference.json``: ``server/smoke_test.py`` compares the served output with it.
"""