Summary

This branch introduces TableBufferPlanner, a table-first buffer allocation planner for deck.gl attributes, and folds the planner-backed buffer orchestration into AttributeManager.

Compared to master, the main change is that deck.gl can now reason about attributes as binary table columns first, then plan how those columns should be represented as GPU vertex/storage buffers under WebGPU/WebGL constraints.

Goals

• Make WebGPU attribute binding more robust by planning vertex-buffer usage explicitly.
• Keep CPU-side binary table columns columnar on the GPU where possible.
• Support constant attributes on WebGPU by materializing them as small real buffers.
• Reduce pressure on WebGPU’s limited vertex-buffer slots.
• Preserve deck.gl table semantics across both:
  • shared instanced geometry
  • row-generated variable-length geometry
• Keep the planner abstract: it works on column descriptors, not Attribute, DataColumn, or AttributeManager.

Highlight: TableBufferPlanner

Adds modules/core/src/lib/attribute/table-buffer-planner.ts.

TableBufferPlanner takes abstract TableColumnDescriptor inputs and returns a TableBufferPlan describing:

• allocation groups
• per-column model binding mappings
• packed vertex-buffer columns
• storage-buffer candidate columns
• unmanaged columns that must keep their existing allocation path

Planner modes:

• table-with-shared-geometry: one shared geometry rendered once for each table row
• table-with-row-geometries: each table row expands into a variable number of geometry vertices

Planner group kinds include:

• interleaved-shared-geometry-columns
• position-attribute-columns
• interleaved-constant-attribute-columns
• separate-attribute-column
• interleaved-attribute-columns
• separate-storage-column
• stacked-storage-columns
• unmanaged-attribute-column

The planner also handles fp64 position high/low mappings, generated row lookup columns, optional storage-buffer planning, and device limit validation.

Core Changes

• Integrates planner-backed allocation into AttributeManager.
• Adds model binding APIs on AttributeManager so Layer can apply:
  • vertex buffers
  • constant attributes
  • index buffers
  • buffer layouts
• Removes the separate GroupedAttributeManager path and simplifies Layer back to using AttributeManager.
• Keeps packed-buffer runtime state inside AttributeManager; TableBufferPlan remains allocation metadata only.
• Preserves unmanaged paths for indexed attributes, transitions, external GPU-only buffers, unsupported fp64 cases, and most noAlloc attributes.
• Allows generated CPU-backed picking colors to be planner-managed to reduce vertex-buffer slot pressure.

Layer/Layout Changes

• Updates layer model creation to pass planner hints such as isInstanced and reserved geometry buffer counts.
• Handles delayed geometry attachment paths such as ColumnLayer.
• Marks non-instanced polygon fill models appropriately.
• Keeps external GPU aggregation buffers unmanaged so GPU aggregation paths can continue publishing their buffers directly.

Documentation

Adds docs/developer-guide/custom-layers/table-buffer-planning.md, covering:

• table-first planner model
• TableBufferPlanner usage examples
• TableBufferPlan
• row lookup columns
• position/fp64 mapping behavior
• buffer groups
• optional storage-buffer planning
• WebGPU constraints

Also updates the custom layer docs table of contents.

Tests

Adds/updates focused coverage for:

• planner allocation groups and mode behavior
• constant buffer step modes
• generated rowIndex and pickingColors
• fp64 high/low position mappings
• external GPU-buffer-only attributes
• constrained vertex-buffer counts
• packed overflow buffers
• layer model binding behavior
• polygon/hexagon-related grouped/planner regressions

Verified locally:

• npx tsc -p modules/core/tsconfig.json --noEmit
• npx tsc -p modules/layers/tsconfig.json --noEmit
• npx tsc -p modules/aggregation-layers/tsconfig.json --noEmit
• yarn test modules/core/lib/attribute/attribute-manager modules/core/lib/attribute/attribute-manager-table-buffer modules/core/lib/layer
• yarn test modules/aggregation-layers/hexagon-layer
• yarn test modules/aggregation-layers/heatmap-layer
• git diff --check

Notes / Review Focus

• TableBufferPlanner is intentionally independent of Attribute, DataColumn, and AttributeManager.
• Storage-buffer planning is currently planner output only; model binding does not yet consume storage groups.
• Packed buffer runtime caches remain in AttributeManager because they store actual GPU buffers and CPU upload arrays, not planner metadata.
• This branch is primarily about the attribute/table-buffer planning architecture and the minimal layer hints needed to exercise it.