# W-077 Status

status: in_progress  
task: 多级 LOD 生成、缓存和视口选择

The cache boundary is now implemented without changing the native WASM ABI:

- `buildLODCacheKey` creates a deterministic, path-safe key from object ID,
  source mesh revision, modifier-stack hash, profile hash and REST/POSE state.
- IndexedDB schema v3 adds `lod_manifest`; the StorageWorker can put/get a
  validated manifest and writes transferable `.mesh` bytes to OPFS
  `projects/<id>/cache/lod/<cacheKey>.mesh`. Read, delete and byte-budget
  pruning update `lastAccessAt` and remove the least-recently-used entries.
- `ThreeLODAdapter` selects pre-generated levels from projected pixel height,
  with a configurable hysteresis band, and is updated by the viewport render
  loop without rebuilding the imported scene.
- `generateLOD` evaluates every requested profile independently from the source
  `.blend` in a fresh WASM module instance, returns a validated manifest and
  transferable geometry, and is exposed by the React inspector. A binary
  length-table container round-trips the level buffers before OPFS persistence;
  the React path reads that persisted container back before installing levels
  into the Three.js viewport, while preserving source transforms and materials.
- Browser E2E covers the OPFS bytes, manifest round trip, stable key and
  hysteresis behavior, multi-level generation and binary geometry round trips.

The React open path now lists all `lod_manifest` rows for the project after a
fresh application load, filters them by the authoritative SceneIR revision and
object IDs, reads the OPFS `.mesh` container, validates the level mesh IDs and
installs the cached levels into the Three.js LOD adapter. Applying any native
scene command also removes manifests from older source revisions, so a stale
cache cannot be selected after a mesh or modifier edit. Browser E2E covers
generation, reload, discovery and reuse.

Pose-aware skin joints are still not exported into the LOD manifest. The native
Collapse bridge now performs one Blender BMesh collapse pass at a time, creating
a fresh heap for each pass and recalculating the quadric choice from the current
mesh state. This removes the previous WASM memory failure at intermediate cube
targets; the LOD Worker keeps per-level isolation as a general fault boundary.
An active modifier stack must now have an `EVALUATED` dependency-graph report
before LOD generation. Metadata-only and unresolved stacks return a structured
capability error rather than caching output from an implicit base mesh.
Each isolated LOD profile applies the same skin/shape-key preflight as direct
Decimate, so a valid `skinPolicy` receives native remapping while a missing or
rejecting policy cannot create a partial cache entry.
Full Blender parity still requires pose-aware skin joints and broader Blender
golden coverage, so the task stays `in_progress`.