perf: bake platformer level tiles into chunked images + merge colliders

[markwpearce]

Summary

Investigates examples/platformer performance: the level's static ground was one Image + one RectangleCollider per tile (233 tiles), costing a per-frame SceneObject in the renderer's depth sort and a collider in the compositor's collision checks for every tile regardless of whether it ever moves.

• BGE.TileMap.bakeTileMapImages() / mergeTileColliderRuns() (new, generic engine namespace, src/source/utils/tilemap/): bakes a sparse tile grid into a handful of composited bitmaps (one per occupied chunk, sized/anchored to that chunk's actual tile bounding box — so a tile whose visual offset pushes it past a grid-aligned bucket boundary is never clipped) and merges same-tag collider runs within a row.
• examples/platformer/Level.bs rewritten to use them: colliders 244→24, scene objects 257→34, on the same level layout.
• Engine fix: SceneObject.isPotentiallyOnScreen() now checks a screen-aligned object's full bounding box (via getBoundingPoints()) instead of collapsing to a single anchor point, so a large baked chunk (or any large Image/Rectangle/Circle/Text) isn't wholesale-culled just because its own anchor corner left the frustum while another corner is still on screen. Regression test in SceneObjectImage.spec.bs.
• Game.debugPrintPerfStats(): new opt-in (off by default, zero cost when off) console-only per-frame stats — entity/collider counts and phase timing (UI input / update+collision / entity draw / swap) — independent of, and a faster cadence than, the existing GC print. This is the tool that surfaced everything else in this PR.
• chunkSize capped at 128 in Level.bs: binary search on both the BrightScript Simulator and a real device found that >=192 silently fails to render a screen-aligned Image at all, despite provably-correct content (confirmed via GetByteArray and a tmp:/ PNG round-trip) and the engine's own draw path reporting success — a real, unexplained platform limit, not a bug in this baking logic. Filed as Support multi-frame loading for Room setup (spread heavy onCreate work like tile baking across frames) #211 (multi-frame room loading) for the constructive follow-up rather than chasing the exact native cause further.

Testing

• npm run check (lint + validate + headless Rooibos suite): 1105 passed.
• Verified visually on both the BrightScript Simulator and a real Roku device (sideload + screenshot) — ground renders correctly, character stands flush on top, before/after perf numbers captured via debugPrintPerfStats().

Related

• Fixes rendering for examples/platformer's baked ground tiles (two real bugs found and fixed along the way: a Y-axis inversion in the chunk-to-Image anchor math, and a tile-footprint clipping bug at chunk boundaries — both covered by new/updated unit tests).
• Follow-up filed as Support multi-frame loading for Room setup (spread heavy onCreate work like tile baking across frames) #211.

🤖 Generated with Claude Code