refactor: extract task environments into separate embodichain_tasks package

docs/superpowers/plans/2026-07-07-task-env-refactor.md

@@ -0,0 +1,19 @@
+# Task Environment Decoupling — Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Extract all task envs from `embodichain` into `embodichain_tasks` package, add entry_points discovery, init hooks, and unified launch.
+
+**Architecture:** IsaacLab-style separation — `embodichain_tasks/` is a separate pip-installable package with its own `pyproject.toml`. Core `embodichain` discovers tasks via `importlib.metadata.entry_points(group="embodichain.tasks")`.
+
+**Tech Stack:** Python 3.10+, setuptools, gymnasium, importlib.metadata
+
+## Global Constraints
+
+- All existing env IDs unchanged
+- JSON/YAML config format unchanged
+- `@register_env` API unchanged
+- `gym.make()` workflow unchanged
+- RoboSynChallenge NOT touched this iteration
+
+---