CoFiI2P geometry bridge and patch-index audit

This follows the initializer-only boundary. It builds CPU model-shaped geometry, source lineage and explicit fine-pixel decoding; it does not run a model, establish official numerical parity or change core defaults. All code here is independently implemented, with no official code copied into core.

Pinned-source findings

Reviewed official CoFiI2P commit ed90edf51cca540ee76152d91133c7e79b63256c:

Source	Observed behavior	Adapter consequence
data/kitti.py:148–180	0.1m voxel averages, then count selection; scarce clouds are repeated	A voxel is a centroid with multiple original contributors, not an original point ID
model/kpconv/preprocess_data.py:56–101	Five levels, half-count sampling with replacement, 128 nearest neighbors	Preserve duplicate rows and every parent index; neighbor matrices index support rows
data/kitti.py:305–327	Nominal 0.5 intrinsics, center crop; fine K scales by 0.5, coarse by 0.125	Feature-grid pixels and source image pixels are different coordinates
model/network.py:156–158,204–225	Centers are (u,v); patches index top:bottom, left:right, then flatten (h,w)	Flat patch index j means column j%4, row j//4
evaluation/eval_all.py:102–105	Adds j//4 to u, j%4 to v	Off-diagonal patch positions are transposed relative to array layout

The patch discrepancy affects 12 of the 16 possible positions, not an observed 12/16 model error rate. We inspected source; we did not run the external model or measure its practical accuracy. The bridge requires an explicit decoding choice: row_major follows image-array layout, and official_eval_legacy reproduces the evaluation formula. Neither is selected using calibration reference errors. Freeze the pilot convention before inference; any legacy comparison must be declared in advance and must not be reported as additional independent data.

The loader reads already prepared image .npy files. crop_original_top_rows=50 in options does not itself establish how those files were made. Color order and upstream top-crop provenance remain unresolved. Existing boundary BGR, top crop and actual-resize pixel-center mapping are explicit local conventions, not a claim that those arrays reproduce a particular checkpoint's training distribution.

Bridge contract

PYTHONPATH=src:. python tools/cofii2p_model_bridge.py prepared/manifest.json options.json bridge-output

options.json requires an explicit seed. Defaults are voxel size0.1m, count20480, five stages and128 neighbors. Input manifest and artifact hashes, options, driver hash, git commit and NumPy/SciPy versions are recorded in the bridge manifest. The existing prepared artifact digest is verified before decoding. Output directories cannot be overwritten.

Voxelization uses a grid anchored at min(initial-camera points)-voxel_size/2, with lexicographically ordered keys. Points, intensity and rotated normals are averaged; averaged normals are intentionally not renormalized. The stored LiDAR centroid uses exactly the same contributor rows as the initial-camera centroid. Thus the two means retain the declared initializer relationship. Never assign an arbitrary contributor's original ID or position to a centroid correspondence.

Lineage is explicit:

level row -> voxel_index_level -> voxel_member_offsets slice
          -> voxel_member_source_rows -> source_point_id

count_selection and parent_level preserve all count/half-sampling choices, including duplicates. points_level/feats are float32. neighbors_level queries the same level, subsampling_level queries the next level against the current one, and upsampling_level queries the current against the next. These are KNN indices, not radius neighborhoods. Use voxel LiDAR centroids for future point–pixel exports and keep each correspondence identity distinct from source IDs.

Fine decoding returns (u,v) on the256×80 feature grid. To match the declared local K convention, multiply fine coordinates by2, then invert the stored source-to-input affine. No additional half-pixel feature offset is silently introduced. Learned receptive-field alignment still needs checkpoint verification.

This SciPy geometry implementation is not a validated CPU inference port. Open3D voxel iteration order and equal-distance KNN ordering may differ. The random stream is locally seeded; it does not reproduce the official dataset-index seed plus intervening random pose draws. Voxelization in initializer coordinates also cannot reproduce the official truth-prealigned cloud's grid in general. These differences require a separate parity experiment before claiming official reproducibility.

Frozen synthetic smoke protocol

Run the committed driver into a new directory:

PYTHONPATH=src:. OPENBLAS_NUM_THREADS=1 python tools/cofii2p_bridge_smoke.py /workspace/work/cofii2p-bridge-smoke-20261030

Use4096 synthetic points on a0.3m grid, declared nonidentity initializer and camera, explicit top crop50, and seed20261030. Run the bridge twice at the default shape 20480/10240/5120/2560/1280, with128 neighbors. Check array equality, complete source lineage, centroid transform, all neighbor shapes/ranges, and39 selected KNN queries against exhaustive distances (allowing different indices at equal distance). Test all16 patch positions and preserve the12 legacy/layout differences. No model, real frame, calibration fit or reference transform file is involved.

The plan pins implementation, tests, driver, options and official-source hashes before the smoke run. Unit tests also cover multi-point voxel means, cancelling normals, exact count rules including divisible scarce counts, KNN orientation, pixel mapping and digest rejection.

GPU/validated inference runtime, suitable weights and terms, checkpoint image provenance and training overlap remain prerequisites for the spent-data pilot.

Outcome

Implementation frozen at b6d0628e. The smoke result passed all eight checks; both bridge manifests and prepared-input provenance are retained beside it. Binary arrays remain under the pinned workspace paths (about49MB including both outputs); manifests record their SHA256. Repeat-run equality was checked for every array, not for manifests whose output paths differ.

Stage counts were20480/10240/5120/2560/1280. Their distinct source-voxel counts were 4096/3775/2581/1504/811: replacement sampling retains duplicate rows, as intended. All39 selected KNN distance checks passed. Pixel layout and legacy decoding differ at exactly12 of16 patch positions. Related boundary/export/PnP tests passed21 cases. There were zero real frames, model calls or accuracy fits. The plan's documentation hash identifies the pre-run protocol at the frozen commit, before this outcome.

Next priority is checkpoint/preprocessing and runtime readiness, followed by an explicitly frozen inference pilot. Additional geometry refinements alone cannot resolve the missing model evidence or establish SOTA.