feat(wh120): add Dexonomy integration and scoped smoke qualification

[horychen]

What this PR establishes

This PR turns the WH120/Dexonomy work into an auditable non-commercial research pipeline:

1. keep canonical DGN5K/OBJ5K grasp payloads as NPY files;
2. resolve each selected (dataset, template, scene, candidate) unit to one canonical grasp_data payload and its complete scene/mesh dependency closure;
3. qualify explicitly named suites with immutable evidence; and
4. bridge only bounded, explicitly selected units into the separate Magichand MuJoCo + motor-simulation repository.

The accepted provisional project charter and definition of done are now versioned in docs/PROJECT_GOAL.md. Corpus-wide GIF/NPZ generation is explicitly out of scope: previews and dense telemetry are opt-in diagnostics, not a replacement dataset.

Upstream source and license

• project: Dexonomy
• source: JYChen18/Dexonomy
• paper: Dexonomy: Synthesizing All Dexterous Grasp Types in a Grasp Taxonomy
• upstream dataset: JiayiChenPKU/Dexonomy
• upstream work/data license: CC BY-NC 4.0, not MIT

Attribution and the non-commercial restriction are recorded in NOTICE-WH120.md, the README, generated provenance, and the project charter. This PR does not grant commercial rights to Dexonomy-derived code/data or to separately owned WH120/object assets. Commercial use remains blocked pending written permission from all relevant rights holders.

Data facts and scale

• The provider reports that the combined DGN5K + OBJ5K source/assets/results footprint is 60+ GB.
• grasp_data is the canonical saved pre-grasp, grasp, and squeeze joint-angle payload.
• succ_grasp and new_tmpl are link/index views of grasp_data; they are not additional independent angle datasets and must not be summed as unique units.
• A unit is (dataset identity, template, scene identity, candidate), with one canonical NPY plus references to a frozen scene and shared mesh closure.
• Object meshes intentionally live in the shared DGN/Objaverse asset stores. A result bundle without copied meshes is therefore not automatically “missing meshes.”
• The compact index stores paths, hashes, provenance, and link semantics without copying meshes or NPY payloads. Its new-output budget defaults to at most 20 GiB. The existing 60+ GB source corpus itself is not claimed to fit inside 20 GiB.

Audited result packages remain separate identities:

• OBJ5K second round: 0821-2026-second-round-obj5k/obj_train_wh120_full_20260821_1914.tar.gz, 4,682,479,789 bytes, SHA-256 5bc3a9de864c296797a7a74afe6d73a4e2c6061b91b117b3ccc61c5cf21c33cf, producer commit b606bfd.
• DGN5K currently available: 0818-2026-first-round/full_train_wh120_bundle.tar.gz, 6,715,294,762 bytes, SHA-256 c082583e0afd5f6e37c591a6215eb1e0ad14f2e1a1d959520a4cd2b1347f9bf7. It remains labelled first round; this PR does not invent or relabel a DGN second-round package.

Implemented here

• native WH120 hand model, collision/keypoint assets, annotations, configuration, and Ubuntu Blackwell runtime support;
• scoped suite registry for generic-dexonomy31-v1, distal-scissors-v1, tripod-variation-chopsticks-v1, and an explicitly unqualified diagnostic subset;
• strict specialty exact-manifest loader (approved scissors/chopsticks manifests are deliberately not fabricated);
• schema-v4 streaming compact index with canonical link semantics, namespace census, provenance, uniqueness, foreign-key checks, crash recovery, completed-resume byte stability, and fail-closed path/link checks;
• shared transitive scene-asset closure across XML/URDF/OBJ/MTL/texture dependencies, rejecting path escape and every symlink component;
• immutable run freezing and verification, exact source/runtime/config/dataset/payload binding, cooperative locks, terminal replay, campaign-level authority checks, and no-clobber sealing;
• safer full/KP runner contracts: no broad process kills, explicit GPU/input/output contracts, fail-fast cleanup, and preflight mode;
• focused tests for index recovery/concurrency, asset closure, loader semantics, suite policy, run freezing, per-run verification, campaign sealing, and launch-script contracts.

The downstream Magichand bridge is reviewed separately in wuji-technology/wuji-hontai-magichand#5. It validates a real DGN moving-root command through MuJoCo/motor plumbing; it must not be interpreted as generic31/all-33 or physical-hand qualification.

Validation at commit 7a67cf709d5290fd49ecf73008fbb0cf943da471

• exact committed tree: caabbe4aa78d8697b39759b3da9250ad66d2eca1
• full Ubuntu test suite in an isolated environment: 265 passed, 5 skipped
• Windows focused regression: 109 passed, 9 skipped
• Ubuntu exact-chain regression: 109 passed
• git diff --check, shell syntax checks, and uv lock --check: pass
• dual-RTX-5090 preflight: both GPUs discovered, Warp 1.8.1 / CUDA 12.8, one real DGN scene selected from 34,182, no workers launched; run-identity SHA-256 18eea3f9fb8b347f46875e1a4a6ad32841d3183a4a9d871b50a3109efe75de0f
• real schema-v4 DGN one-unit index: 1 canonical grasp, both alias views, 1 scene, 97 mesh records; 180,224 bytes; foreign-key check clean; completed --resume preserved byte-for-byte SHA-256 d9f2fa1aa9331b8f79e1290358e2d9beefa959845e9ee653509c89fae08fb57f
• independent final audit found no remaining P0/P1 code issue in this checkpoint.

What is deliberately not claimed

• 32_Ventral still needs WH120-native human re-annotation and approval. Numerical clamping is not a substitute.
• 19_Distal and 21_Tripod_Variation need dataset-owner-approved, hash-pinned scissors and chopsticks manifests before specialty qualification.
• A newer DGN5K second-round producer package/hand snapshot has not been supplied or verified.
• No qualified generic31 or all-33 campaign is claimed until the three suites are run and independently sealed after those inputs exist.
• MuJoCo constraint-solver wrench is simulated contact evidence, not tactile-sensor ground truth; simulation is not physical-hand or sim-to-real acceptance.
• Dense GIF/telemetry batching is downstream Magichand work and remains bounded/opt-in; this PR does not launch a corpus-wide export.

External approvals still required

• WH120 annotation owner: approve a corrected 32_Ventral.
• Dataset/qualification owner: approve scissors/chopsticks scenes.
• Data producer: provide any newer DGN package and producer hand snapshot.
• Asset/upstream rights holders: clarify redistribution and commercial permissions.
• Hardware validation owner: define and sign off physical-hand acceptance.

This is a runnable and auditable research checkpoint, not a claim of all-33, physical, or commercial readiness.

[horychen]

@FeilaoWu 麻烦 review 这个 WH120/Dexonomy 研究集成 checkpoint，当前提交为 7a67cf709d5290fd49ecf73008fbb0cf943da471。

建议重点看：

• schema-v4 compact index、grasp_data / succ_grasp / new_tmpl 链接语义与 20 GiB 新输出预算；
• scene/mesh 传递闭包、冻结/验证/campaign seal 的 fail-closed 合同；
• 32_Ventral 人工重标 gate，以及 Distal/剪刀、Tripod Variation/筷子 exact-manifest gate；
• Dexonomy 官方出处与 CC BY-NC 4.0 非商用边界。

验证：Ubuntu isolated full suite 265 passed, 5 skipped；双 RTX 5090 real-data preflight 通过；真实 DGN schema-v4 一单元索引完成且 resume SHA-256 字节不变。下游 Magichand bridge 在 wuji-technology/wuji-hontai-magichand#5。

当前账号没有 base repo 的 RequestReviews 权限，因此无法把你设成 GitHub formal reviewer；先在这里直接邀请，上游 maintainer 可补点 reviewer。