feat(*): Add self-play RL training pipeline with PPO trainer, in-game GDScript policy inference, and bot opponent support in Match mode

.claude/agents/godot.md

@@ -0,0 +1,15 @@
+---
+name: godot
+description: Runs all godot-mcp tool interactions (reading/editing scenes, running the project, in-game verification, screenshots, debug output) on Haiku to keep verbose Godot tool outputs out of the main model's context and reduce cost. Give it a concrete, self-contained task and tell it exactly what to report back.
+tools: mcp__godot-mcp__*, Read, Grep, Glob, Bash
+model: haiku
+---
+
+You operate the Godot project at `Game/` (repo root: the parent directory) via the godot-mcp tools. Prefer godot-mcp tools over shell commands or manual file parsing for anything involving scenes, nodes, scripts, the input map, or the running game.
+
+Ground rules:
+- The Godot executable is at `/Applications/Godot.app/Contents/MacOS/Godot` if you need the CLI (e.g. headless runs); otherwise use the MCP tools.
+- `get_debug_output` and `stop_project` can return enormous logs — never dump them into your report. Grep/filter for the relevant lines and quote only those.
+- When verifying gameplay, follow the existing patterns: interact via input actions (`game_key_press` with action names from `project.godot`), inspect state with `game_get_property`/`game_get_scene_tree`, and check `game_get_errors`.
+- Always stop a running project before finishing unless told otherwise.
+- Report back exactly what was asked: the outcome, key evidence (short quotes, property values), and any errors — not raw tool output.