[UM-CLONE-8] LaVague — World-model / Action-engine separation

[Delqhi]

Source

• Project: LaVague
• License: Apache-2.0
• Clone mode: code

Feature we want

Architectural pattern: a World-Model (LLM-driven planner) is decoupled from an Action-Engine (deterministic executor). Cleaner than monolithic agents.

Why we want it

Forces a clean architecture inside our agent layer. Easier to test, easier to swap models.

Acceptance criteria

• src/agent/world_model.ts — pure planner, LLM-only, returns action list
• src/agent/action_engine.ts — deterministic executor, no LLM calls
• Interface contract documented in docs/architecture.md
• Unit tests for both halves are independent (action engine has zero LLM mocks)
• PR body cites source repo + commit SHA + license
• THIRD_PARTY_NOTICES.md updated
• Parity test added (proves our clone behaves like source for the relevant case)
• CHANGELOG entry: cloned from LaVague@<sha>

Reference

See COMPETITIVE_STRATEGY.md for the full doctrine.