Qualify PointWorld on Flat'n'Fold before target access

[FlorianPfaff]

Objective

Execute the source-only qualification frozen in protocols/pointworld-flatnfold-source-qualification-v1.json before any Flat'n'Fold target outcome is opened.

This is the proposed fresh-provider path for issue #49: PointWorld -> Prob4D -> BayesianPhysTwin on a later garment-disjoint Flat'n'Fold robot cohort.

Frozen source revisions

• Prob4D: d02f057671023ff586ebfc15c904dbb0a60f4425
• BayesianPhysTwin: b5f07d649ac2cd7dc6ca1aceb4004ff4803bddc8
• PointWorld: 05484826dfef74cbe278a3974179a5a16705d35d
• Flat'n'Fold code: fa0d3d17ac827e7b5c43ec1ef7c0c38ad5e39340

Representation gate — software decision completed in PR #334

PointWorld predicts seeded scene-point trajectories/flows on B x T x N_scene x 3 with a B x T x N_scene x 1 log-variance output. The existing dense PredictionWindow v2 contract is not used.

PR #334 implements and freezes sparse-persistent-point-contract-v1:

• persistent integer point identities and seed positions;
• seed-relative displacement trajectories;
• separate visual-source and planned-action lineage;
• explicit dependence groups;
• raw PointWorld log variance retained without a metric-covariance claim;
• no dense rasterization, target-truth nearest-neighbour mapping, or target-tuned interpolation.

The machine-readable decision is protocols/pointworld-sparse-representation-v1.json; implementation and tests are under pointworld_sparse_prediction.py, pointworld_sparse_adapter.py, and test_pointworld_sparse_prediction.py.

Source-only checklist

• Bind exact PointWorld checkpoint SHA-256 and runtime environment.
• Bind exact PointWorld normalization-statistics SHA-256.
• Bind exact Flat'n'Fold dataset-byte manifest for the inventory/source cohort.
• Freeze complete garment identities before provider residuals.
• Verify three-camera timestamps, intrinsics, and extrinsics.
• Verify Baxter action-pose semantics and construct PointWorld robot point-flow actions without target state.
• Verify metric coordinate transformations using released calibration only.
• Verify PointWorld scene-point identity semantics across the ten-frame forecast horizon on a source sample.
• Resolve the software representation gate without inventing a dense image topology.
• Build an outcome-blind ProviderSupportFeasibilityV1 request before prediction payloads, residuals, or target outcomes are opened.
• Stop and retain a valid negative if support fails.

Authorization rule

Only after every remaining source-side requirement is frozen and support-positive may a new held-out protocol version bind source/calibration/target garment rosters, primary provider arms, one BayesianPhysTwin query, clustered inference, and numerical promotion gates.

No checkpoint, representation mapping, mask, split, exclusion, covariance rule, or guard may be retuned after target access.

Claim boundary

This issue can establish source feasibility only. The implemented sparse contract resolves a software representation question, not PointWorld competence, Prob4D accuracy or calibration benefit, BayesianPhysTwin benefit, Causal4D benefit, or deployment safety.

[FlorianPfaff]

Representation sub-gate implementation is now in draft PR #336.

Source-only finding retained by the PR:

• PointWorld samples scene indices at the first timestep and applies the same selected indices across its complete forecast horizon;
• its exported scene_flows are absolute trajectories for those seeded window-local points;
• Prob4D therefore gains a strict sparse persistent-point payload and prediction import-pointworld-sparse route instead of target-dependent dense rasterization;
• PointWorld raw log_var remains explicitly uncalibrated provider uncertainty.

The PR does not authorize target access. Exact checkpoint/runtime identity, Flat'n'Fold bytes and garment roster, Baxter action conversion, camera/metric-frame qualification, source runtime export, and ProviderSupportFeasibilityV1 remain open under this issue.

[FlorianPfaff]

Source-asset staging update (2026-08-29): gpuserver4090 now has a persistent smoke-test root at /home/github-runner/.cache/datasets/pointworld-flatnfold-smoke-v1 on FireCuda, readable by github-runner and other users. Present and verified: PointWorld large-droid+behavior checkpoint (13.04 GB; SHA-256 verified), pinned PointWorld code, DINOv3 code, PointWorld normalization statistics, and Flat'n'Fold Hardware/calibration files. Two external acquisition blockers remain before the first real source-only smoke run: (1) approved DINOv3 model weights from Meta; (2) one complete Flat'n'Fold robot demonstration, because the previous dataset link expired and access has been re-requested. No target garment outcome, provider residual, covariance result, or BayesianPhysTwin innovation has been opened. While those two assets are pending, the next software step is to freeze/verify the local asset manifest and runner-path contract, and finish the source-only Baxter/gripper action-to-PointWorld action-flow adapter so that the first complete demonstration can be run immediately once available.