Deduplicate CAD assets and stop re-fingerprinting undo history during .skf autosave

[davidsteinmeyer]

Summary

Autosave generates a full .skf archive containing the complete undo history. Export was encoding and SHA-256-hashing every resource in every history state, deduplicating only after hashing, inlining the large cadDisplayEdges cache into every node of every state, and re-fingerprinting the whole history twice. On a real 6.9 MB project this produced a 30.8 MB project.json and a ~3.8 s export on the main thread. This PR makes asset handling content-addressed before the expensive work, stores display edges once as a shared asset, and reuses history fingerprints. History behavior is fully preserved.

Root cause

• SkfArchiveBuilder.addAsset computed sha256Hex(bytes) before checking its kind:sha256 dedup map, so identical resources were re-encoded and re-hashed once per state. addDerivedMesh cached only by object identity, and cadBrep, brepStep, and image data URLs had no pre-hash cache at all.
• cadDisplayEdges was left in the node definition and inlined per node per state. In the real project 230 occurrences held only 2 unique payloads (0.21 MB unique -> 24.5 MB in project.json).
• canonicalizeShape always allocated new nested arrays, so the fingerprint WeakMap missed, and exportSkfProject fingerprinted the entire history twice.

Changes

• Added recordByResourceKey and addEncodedAsset(kind, resourceKey, encode, options) to SkfArchiveBuilder; it returns a cached promise on a key hit and only invokes the encode() thunk on a miss, so encoding and hashing are skipped for repeated content.
• addDerivedMesh now keys on mesh.storageResourceId ?? immutableResourceFingerprint(mesh) (content-based) instead of object identity.
• serializeShapeNode routes imagePlate, sketchProfile.images, brepStep, and cadBrep through addEncodedAsset with content keys.
• cadDisplayEdges is now extracted from the node definition, stored once in the asset table (kind brep, media type application/vnd.sketchforge.display-edges), and referenced by the new optional cadDisplayEdgesAssetId node field. The importer restores it and also shares one array per distinct payload for legacy inline packages.
• exportSkfProject reuses the fingerprint already computed by hydrateEditorHistoryState and only re-fingerprints entries whose shapes were actually repaired.
• repairDuplicateGroupedObjectIds returns the original array when nothing changed.
• canonicalizeShape is now identity-preserving when a shape is already canonical, so resource fingerprints cached by object identity survive repeated canonicalization.

Why this fixes the issue

• Hashing drops from O(history_states x resources) to O(unique_resources); on a 200-state shared-resource test, digest calls fall from 201 to 2.
• cadDisplayEdges storage drops from 24.5 MB to 0.21 MB, shrinking the real project's project.json from 30.8 MB to ~6.5 MB and the archive from 7.8 MB to 1.9 MB.
• History fingerprinting is now cached and single-pass.

Testing

• npm run typecheck
• Manually tested the affected workflow

Notes

New unit tests:

• "encodes and hashes each unique history resource once instead of once per undo state"
• "deduplicates large display-edge caches across undo states"

Performance impact

• Real project autosave: ~3.8 s -> ~1.0 s; archive 7.8 MB -> 1.9 MB.
• History fingerprinting: ~1.4 s -> ~0.1 s.
• Shared-resource synthetic case (200 states): 201 -> 2 crypto digests.