RFC: propose @deck.gl-community/cosmos-layers

[ibgreen]

Summary

• adds an RFC for a new opt-in @deck.gl-community/cosmos-layers package backed by @cosmos.gl/graph
• evaluates a CPU-readback CosmosLayout prototype and a shared-device, zero-copy CosmosGraphLayer
• recommends the shared-device design as the production architecture
• documents lifecycle, coordinate-system, testing, benchmark, release, and maintenance requirements
• adds the RFC to the documentation sidebar

Motivation

cosmos.gl provides a compelling GPU force simulation, but its current high-level Graph API owns the canvas, render loop, render passes, interactions, and synchronous position readback. Wrapping that API directly would either create a second-canvas overlay that is not a true deck.gl layer or lose the main performance benefit through per-tick GPU-to-CPU copies.

The RFC establishes a package boundary and a phased path before implementation begins.

Proposed direction

1. Keep the dependency isolated in a separate cosmos-layers package.
2. Use an experimental, throttled readback adapter only for semantic validation and benchmarks.
3. Require shared-device, externally scheduled, zero-copy operation for a production CosmosGraphLayer.
4. Coordinate upstream cosmos.gl improvements, including a simulation-only API, optional DOM/canvas ownership, external scheduling, read-only GPU position resources, host render-pass support, efficient snapshots, indexed point pinning, luma.gl dependency alignment, and backend capability reporting.

Impact

This PR is documentation-only. It does not add dependencies, runtime code, or a published package. The open questions in the RFC are intended to guide deck.gl-community and cosmos.gl maintainer discussion.

Validation

• yarn lint
• yarn test-node — 168 files passed, 1,417 tests passed, 9 skipped
• cd website && yarn build

The website build completed successfully with existing dependency-age and CSS-minifier warnings.

[github-actions]

Coverage

Stmts: 51.01% | Branch: 42.11% | Funcs: 49.15% | Lines: 51.12%

Test Files 168 passed (168)

[ibgreen]

@rokotyan @Stukova