Pinned CPU checkpoint execution and raw correspondence export

This extends the random-weight CPU forward probe into an explicit checkpoint-loading boundary. The user has no GPU; this path stays on CPU. Official CoFiI2P weights are still absent locally and their README distribution hosts remain excluded by the enforced network policy. No repeat denial probes, unsupported proxy routes or speculative mirror downloads were performed.

Request and loading contract

PYTHONPATH=src:. OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=2 \
python tools/cofii2p_checkpoint_cpu.py request.json new-output-directory

The request schema requires a pinned bridge manifest, pinned checkpoint, reviewed source checkout, checkpoint kind/source/training/terms declarations, image-channel policy, patch layout and seed. Unknown fields, including reference-problem inputs, are rejected. Metadata is a caller declaration, not automatic verification of official weights, license or training isolation. Unknown terms/training must remain explicit.

Checkpoint and input digests are verified before model construction. The prepared artifact/manifest chain is checked and the bridge image must equal the prepared image. No real problem/reference loader, dataset calibration or labels enter the forward call. The runner uses the already exercised isolated CPU source patch; no external model code enters Calibrex core.

Load raw tensor state only with torch.load(map_location='cpu', weights_only=True). Require exactly the model's keys, shapes and dtypes, and finite CPU tensors, then load with strict=True. No module. prefix stripping, wrapper guessing, silent dtype conversion, unsafe-pickle fallback, partial load or random-weight fallback. If any tensor fails validation, the model state is not mutated. This matches the pinned official training code's raw model.state_dict() save format; a differently packaged checkpoint requires an explicit, separately reviewed conversion.

Correspondence coordinates and source lineage

An observational wrapper records the native fine_process selected coarse rows; it does not modify returned features, scores or indices. Forward receives None for all label arguments. Repeat-forward outputs and selected rows must match exactly, and every returned point is checked against its selected input row.

Fine-patch cosine matching uses native PyTorch normalization and first argmax. The request explicitly chooses row-major or legacy evaluation pixel decoding; the prior patch audit still applies. Fine-grid pixels map through the declared local affine to original-image pixels. The input image policy is either the prepared BGR array or an explicit BGR-to-RGB swap. Neither choice establishes official checkpoint preprocessing parity.

raw_pairs.npz contains LiDAR-frame voxel centroids, original-image pixels, unique correspondence IDs, selected coarse rows/voxel indices, fine pixels/patch indices and CSR arrays containing every contributing original point ID. Duplicate centroids are retained as separate predicted pairs. No nearest-source snapping, invented covariance/confidence, correctness filter or pose fit is added.

The result schema records checkpoint/input/source hashes, loaded-state signature, runtime, timings, memory, repeatability, raw pair artifact and all failures. CLI exits2 on a recorded execution failure; validation failures exit nonzero before inference. Raw exports require a separately declared provider/uncertainty/quality policy before any probabilistic calibration claim. Pair count alone is not acceptance or accuracy.

Frozen integration test

The official learned artifact cannot be tested yet. Instead, the committed fixture builder constructs the full pinned model once at seed20261030 and saves its raw state as SYNTHETIC_RANDOM_STATE.pt. It is never labeled as trained or official. The loader refuses fixture execution unless --allow-fixture is explicitly given. Both checkpoint creation and execution occur in separate fresh processes.

Use the existing standard20480-point synthetic bridge and its actual prepared constant BGR image. Fix row-major decoding and two CPU threads. Run once with a180s ceiling; retain failure if random features produce insufficient pairs. Do not swap seeds/images or manufacture matches to make this fixture succeed. Checkpoint tensor loading is independently unit-tested with rejection of wrong keys/shapes/dtypes, nonfinite/non-tensor values and nested wrappers, with state immutability on failure.

The plan pins the implementation, fixture builder, tests and bridge before the integration run. This is a software integration check, not model accuracy, a calibration fit or an official-weight run. No learned dataset or fresh holdout is downloaded. Next obtain an official KITTI artifact through an authorized accessible location, establish checkpoint/image/ training provenance and run the same CPU loader before freezing the spent-data pilot.

Integration outcome and learned-weight handoff

Frozen implementation 5b90e2e6. The separate-process full-model fixture run completed: result, request, raw pairs. Strict CPU loading restored the raw state, then the actual prepared-image/test path completed with exact repeat outputs/selected rows. Initialization plus load took5.82s, first forward10.21s, peak process RSS1.92GiB, on two CPU threads. Timing excludes geometry preparation/source copying and is specific to this untrained fixture, not a learned-model performance guarantee.

All4 raw pairs were retained, including duplicates:3 distinct voxel centroids and 3 distinct image pixels. This is not sufficient independent support for a pose acceptance claim, and no pose was fitted. Centroids and all contributor IDs were independently reconstructed from the original prepared points; mapping source pixels forward reproduced the fine-grid coordinates. Related boundary/export/PnP tests passed32 cases. Request/result schemas and frozen source/input/output hashes were validated. The plan's document hash refers to the pre-run protocol.

The196.97MiB fixture checkpoint stays outside git in the recorded workspace path; it is random initialization, not an acquired learned artifact. It uses the same initialization seed as the target process; loader mutation behavior is separately verified with different source/target states in the small-model unit test. Do not infer learned recovery or generalization from this integration result.

For a real artifact, fill the schema-valid learned request template with actual paths/SHA256 and source/training/terms declarations, then run the CLI without --allow-fixture. Zero hashes and placeholder paths deliberately cannot pass artifact validation. Template image/layout choices illustrate the local contract; they do not settle official checkpoint preprocessing. Declare and freeze the intended policy before any real-data comparison. Official weights remain unavailable; this completes the executable loading/export boundary only.