Changelog

Highlights from the engine changelog, grouped by area and linked to the docs. The current workspace version is 1.1.0 (Cargo.toml).

​

1.1.0 — current

​

Unified Editor & Data-Driven Levels

• euca-studio — one unified editor. A single app that discovers every project, lets you select one in-editor, edit its scene as data, and press Play to run it. cargo run -p euca-studio. The game-engine / experience face. →
• Headless host. cargo run -p euca-studio -- --headless [project_id] serves a project’s world over the agent HTTP API on :3917 with no window — it runs the game’s full per-tick logic (game.step), replacing the old fixed-schedule example server. →
• Data-driven levels via generic reflection. Any #[derive(Reflect)] component survives a save↔load round-trip, so the host loads a project’s level and the editor edits the same scene the game plays.

​

World-model & evaluation substrate

• euca-online — interactive world substrate. OnlineWorld exposes observe / peek / step, with the tested peek == step invariant.
• Regret grading. Prequential regret in nats against the known generative distribution; a perfect predictor scores 0.
• fork_peek counterfactual branching — clone mid-episode; the unchanged remainder stays bit-identical. →
• euca-dataset — world-model data engine. Object-centric state/action/event extraction consuming ground-truth segmentation/depth channels. →
• Editor Evaluation mode — a ground-truth inspector that makes observe/peek/fork/regret visible.

​

Unified rule engine

• euca-rule — pure-data Rule DSL. One Rule (condition + declared distribution + effects) drives MOBA, agent, editor, and dataset paths. →
• Single-RNG CI gate — one deterministic RNG source across the rule/online path.

​

Physics

• Iterated SI solver + rotation-aware narrow-phase, convex colliders, static triangle meshes, compound colliders via convex decomposition, and reduced-coordinate articulation (Phases 1–4). →

​

ECS scale

• Chunked parallel queries — the 5-system pipeline runs ~11.3× faster at 1M entities (2.7 ms vs 30.6 ms on M4 Pro). spawn_batch (~16×). Commands deferred mutation. →

​

Rendering

• Velocity pass wired (TAA + motion blur no longer no-ops), real per-frame frustum culling + LOD, a post-process Sobel outline, and GPU-timing in the profiler. →

​

Asset pipeline

• AI-3D-gen world-model decoupling + provenance — generated appearance provably never touches the hashed state_digest. Cooked-asset LOD pipeline with PNG-compressed textures. →

​

RHI & native Metal

• euca-rhi — RenderDevice trait decoupling the renderer from any GPU API, plus a native Metal backend (Apple Silicon).

​

Foundations

The base the above builds on: the build/play/experiment HTTP surface, the gameplay layer (with MOBA and a GTA-like vertical slice shipping as example games), the forward PBR renderer, and the physics solver.