Refactor/skeleton frontend

[seankmartin]

PR Summary

Refactor: Align skeleton render layer with segmentation layer patterns

This PR refactors src/skeleton/frontend.ts and related files to bring the skeleton rendering architecture in line with the more mature segmentation render layer patterns — cleaner interface boundaries, shared GPU utilities, reactive shader parameters, and elimination of duplicated pick-handling and rendering logic.

What changed

Interface modernization — Old mixed-concern interfaces (SkeletonLayerInterface, SkeletonChunkInterface) replaced with clearly separated types:

• SkeletonShaderContext — rendering context (gl, displayState, shader parameters)
• SkeletonShaderParameters — shader feature flags as a WatchableValue (replaces scattered booleans)
• SkeletonGPUGeometry — pure GPU geometry interface
• PackedSkeletonGeometry — CPU-side vertex/index data

GPU upload consolidation — gpu_upload_utils.ts deleted; its core logic was already in texture_access.ts. Added updateOneDimensionalTextureElement() for in-place single-element texture updates, and extracted uploadSkeletonChunkToGPU() as a shared helper for both chunk types.

SkeletonOverlayChunk promoted to full class — Now implements SkeletonGPUGeometry, owns its GPU resources (textures, buffers, picking data), and exposes updateSelectedNode() which patches a single texture element in-place instead of rebuilding all GPU state when only node selection changes.

Overlay geometry memory optimization — Introduced a reusable scratch buffer for arrays that are only needed during GPU upload (segmentIds, selected, edgeIndices, edgeSegIds), avoiding repeated allocation. Arrays needed for CPU-side picking (positions, nodeIds, pickSegmentIds) are still freshly allocated and owned by the chunk.

Two-level overlay caching — Geometry key excludes selectedNodeId, so selection-only changes trigger the fast updateSelectedNode() path. Geometry changes trigger a full rebuild. A frame-number guard prevents redundant work across multi-pass rendering.

Shader parameter management — skeletonShaderParameters as a WatchableValue<SkeletonShaderParameters> replaces scattered booleans. Reactive update registration listens to segmentStatedColors and segmentDefaultColor, decoupling shader feature flags from layer state.

Picking logic deduplicated — Both PerspectiveViewSpatiallyIndexedSkeletonLayer and SliceViewPanelSpatiallyIndexedSkeletonLayer had verbatim copies (~75 lines each) of transformPickedValue and updateMouseState. Extracted as free functions transformSpatiallyIndexedSkeletonPickedValue and updateSpatiallyIndexedSkeletonMouseState; each class method now delegates in 1–4 lines.

Cast cleanup — as SkeletonLayerDisplayState & SpatialSkeletonDisplayState intersection casts removed throughout draw(), isReady(), and constructors. chunkGeometryRenderLayerInterface renamed to browsePassLayerView.

Naming and removals — SkeletonRenderMode enum deleted (render_mode.ts removed); SpatiallyIndexedSkeletonParameterAccess union type removed; DEBUG_SPATIAL_SKELETON_CHUNKS set to false.

---

What's Left to Refactor

High priority

Segment ID precision bug (frontend.ts ~L625)
The existing TODO captures this: segmentAttribute is Uint32 but segments are Uint64. Fixing this also unblocks the structural cleanup below — once the attribute is known to be internal, it can be removed from vertexAttributes entirely.

Internal attributes mixed into vertexAttributes (frontend.ts ~L625, ~L630)
segmentAttribute and selectedNodeAttribute live in the same vertexAttributes array as user-defined attributes but are skipped by index check in both shader getters:

if (i === this.segmentAttributeIndex || i === this.selectedNodeAttributeIndex) continue;

These are internal implementation details, not user-defined. They should be stored separately so the user attribute loop needs no special-casing and the boundary is explicit. This also makes the code clearer about what a user can address in the shader vs. what is managed internally.

Medium priority

Edge and node shader getters share ~80% of their body (frontend.ts L516–860)
Both edgeShaderGetter and nodeShaderGetter call parameterizedEmitterDependentShaderGetter with identical config (parameters, fallback, extraParameters, shaderError), then inside defineShader:

• Same preamble: defineCommonShader, defineAttributeAccess, defineDynamicSegmentAppearance
• Same three-way fragment color branch (dynamic appearance / legacy uniform / per-vertex attribute)
• Same segmentColor() definition logic
• Verbatim identical user attribute varying loop at the end

The only genuine differences are the primitive-specific parts: edge uses defineLineShader + paired index vertex logic; node uses defineCircleShader + gl_InstanceID + selected-node outline. The shared preamble, color branch, and user attribute loop could be extracted into helpers — e.g. defineShaderPreamble(), defineSegmentColorFragment(), defineUserAttributeVaryings() — so each getter only contains the primitive-specific parts.

attach() bodies are identical (frontend.ts ~L3229, ~L3500)
Both view classes have ~30 lines of identical attach() logic; the only difference is "3d" vs "2d" passed to getSources(). Same situation as updateMouseState was before this PR — can become a shared free function attachSpatiallyIndexedSkeletonLayer(base, self, attachment, view: "2d" | "3d").

Low priority / cleanup

• data: any in updateMouseState — Both view classes still have data: any; should be unknown to match the direction of removing casts.
• getValueAt unused param convention — position should be _position to match the _pickedValue convention already in the same methods.
• registerDisposer(base.redrawNeeded...) placement — Registered near the top of the 3d constructor, at the bottom of the 2d constructor; should be consistent.
• PackedSkeletonGeometry.lod field (~L152) — Unused, should be deleted.
• Scratch buffer shrinking (overlay_geometry.ts) — gpuScratchBuffer grows monotonically; needs a trim/clear strategy under sustained use.