fix(plugin-trees): consume dirty marks so scene-ready fires for plant scenes

[wass08]

What does this PR do?

Instanced plant kinds (trees:tree, trees:grass, trees:flower) never cleared their scene-store dirty marks. FloorElevationSystem deliberately leaves the mark for any kind with a def.system ("the downstream consumer will clear it"), but InstancedKindSystem never participated in the dirty protocol — so the marks lived forever, hasPendingSceneBuildWork() never went false, and the Viewer's scene-ready signal stalled at SCENE_READY_MAX_WAIT_FRAMES (180 frames) on every scene containing a plant.

The visible cost: on the headless bake worker (SwiftShader, ~1s frames) every plant-containing scene paid ~190s of pure cap-wait per bake — measured as pageSettleMs ≈ 193s on prod bakes of a plant-heavy scene, regardless of warm caches.

The fix adds a priority-2 useFrame pass to InstancedKindSystem that clears this kind's dirty marks:

• priority 2 — runs after FloorElevationSystem's priority-1 lift pass in the same frame, so the mark survives long enough for the lift to apply (same contract GeometrySystem/ItemSystem rely on);
• only for registered nodes — a node whose selection proxy isn't in sceneRegistry yet keeps its mark for a later frame;
• instances rebuild synchronously from the store (the useMemo chain), so a rendered node is already "built" — there is no async work the mark needs to outlive.

How to test

1. bun dev, open a scene containing trees/grass — verify plants render, hover/selection outlines work, and moving a plant on an elevated slab still sits at the right height.
2. Bake a plant-heavy scene headlessly (bun apps/community/scripts/bake.ts <id> from private-editor) and check the [bake:timing] line: pageSettleMs should be actual load time (~seconds), not ~180 × frame-time.
3. Export correctness: baking the same scene with and without this change produces a byte-identical GLB (verified on a 35 MB plant-heavy prod scene — 35,193,812 bytes both ways; settle dropped 22.5s → 11.6s locally).

Screenshots / screen recording

N/A — non-visual change (render output and exported GLB verified byte-identical; only the readiness signal timing changes).

Checklist

• I've tested this locally with bun dev
• My code follows the existing code style (run bun check to verify)
• I've updated relevant documentation (if applicable)
• This PR targets the main branch

🤖 Generated with Claude Code

---

Note

Low Risk
Small, localized change to dirty-mark lifecycle timing; no auth/data paths and render output is intended to stay identical.

Overview
Instanced plant kinds (trees:tree, grass, flowers) now participate in the scene dirty-node protocol by clearing their own marks after each frame, instead of leaving them set indefinitely.

InstancedKindSystem gains a priority-2 useFrame pass that calls clearDirty for dirty nodes of this kind that are already in sceneRegistry. That runs after FloorElevationSystem’s priority-1 lift (same ordering other def.system consumers use). Nodes without a mounted proxy keep their mark until registration catches up.

Effect: hasPendingSceneBuildWork() can go false when only plants were dirty, so the viewer scene-ready signal and headless bakes no longer sit at the ~180-frame cap; rendering is unchanged because instances already rebuild synchronously from the store.

^{Reviewed by Cursor Bugbot for commit 93bcdae. Bugbot is set up for automated code reviews on this repo. Configure here.}