M03-RND-TERRAIN-CHUNK-QUEUE-001 — Pure terrain chunk geometry queue planner

Ownership

Extract chunk-geometry create/remove desired-operation calculation from the streamer into a scene-free TerrainChunkGeometryQueuePlanner.

Exclusive: src/render/terrain/terrain_chunk_geometry_queue_planner.gd, src/tools/verify_terrain_chunk_geometry_queue_planner.gd, docs/modules/terrain-chunk-geometry-queue-planner.md, this claim
Shared/hotspots: src/scenes/streaming/streaming_world_loader.gd, docs/modules/world-renderer.md, docs/modules/README.md, RENDER.md, targets/03-renderer-facade.md
Generated/ignored: .godot, native DLL, generated ADT resources, caches and proprietary renderer corpus

Input: loader-owned tile-state read model plus immutable GodotWorldPosition
Output: fresh create/remove request arrays; planner retains no queue state
Removal: current chunks absent from desired state, or all current chunks while tile LOD is desired
Creation: missing or LOD-mismatched chunks, sorted by origin XZ distance
Queue/cache/schema/coordinate versions: unchanged

Requires: integrated terrain LOD planner master b568768
Blocks: broader terrain service extraction and dependency/performance acceptance
External state: none; tests use synthetic tile/chunk dictionaries

Commands: dedicated removal/create/mismatch/tile-fallback/sparse/order and bounded timing verifier; loader source boundary; facade/cache/LOD/internal-access/ manifest/shutdown/scheduler/streaming planner and repository gates
Fidelity evidence: extracted branch/order/distance rules and loader drain order
Performance budget: bounded linear planning plus existing create sort; no engine resource, task, cache or persistent allocation framework

Inline planner API documentation
New module spec with inputs/outputs, data-flow, lifecycle, sequence, ownership, failure, performance and source map
World renderer/module registry/RENDER.md updates

Important name: TerrainChunkGeometryQueuePlanner
Simplest solution: one stateless planner returning two fresh arrays
Rejected complexity: generic operation graph, queue class, callback strategy, spatial index or ownership transfer
Unavoidable complexity: raw dictionaries remain at the loader/ADT adapter boundary
Measured optimization evidence: bounded synthetic multi-tile verifier

State: handoff
Done: pure planner, loader queue adoption, semantic/timing regressions and docs
Next: milestone integrator merge and asset-backed performance acceptance later in M03
Blocked by:

Commit: f800e11 (refactor(M03): extract terrain chunk queue planner)
Results: dedicated contract PASS (cases=8, iterations=100, 235.096 ms); LOD/cache/facade/internal-access/manifest/shutdown/scheduler/streaming/focus/ ground/terrain-query/coordinate regressions PASS; internal inventory remains 42; editor parse has no script errors; documentation/coordination/diff gates PASS; checkpoint dry-run planned 7/7 entries with expected missing-private- corpus diagnostics
Remaining risks: planner adds a second linear tile-state pass; asset-backed p95/p99 and visual acceptance remain pending; queue drains, terrain state and Mesh/Node/RID finalization remain loader-owned
Documentation updated: terrain-chunk-geometry-queue-planner.md, module registry, world-renderer.md, RENDER.md, inline API documentation and data-flow/lifecycle/sequence diagrams