docs: Add evolved architecture diagram with three-phase pipeline

[abrichr]

Summary

• Adds a comprehensive evolved architecture diagram to README.md and docs/architecture-evolution.md
• Modernizes terminology to align with GUI agent research literature
• Shows implementation status with solid (implemented) vs dashed (future) styling

Key Changes

Three-Phase Pipeline

1. DEMONSTRATE (Observation Collection) - capture, privacy
2. LEARN (Policy Acquisition) - retrieval, ML, fine-tuning
3. EXECUTE (Agent Deployment) - grounding, evals, safety gate

Core Concepts Visualized

• Demo-conditioned prompting: The core innovation - "show, don't tell"
• Policy/Grounding separation: Policy decides what, Grounding determines where
• Safety Gate: Runtime validation layer before action execution
• Multi-source data ingestion: Human demos, synthetic data, benchmarks
• Evaluation-driven feedback loops: Success traces become new training data

The Abstraction Ladder

Literal -> Symbolic -> Template -> Semantic -> Goal

Progressive generalization from raw events to goal-level automation.

Retrieval in BOTH Training AND Evaluation

Demonstrates how demo conditioning improves accuracy from 33% to 100% on first-action benchmarks.

Test plan

• Verify Mermaid diagrams render correctly on GitHub
• Confirm terminology aligns with docs/architecture-evolution.md
• Check all links work properly

---

Generated with Claude Code

[abrichr]

Rendered Architecture Diagram

Three-Phase Pipeline:

1. DEMONSTRATE - Capture human demos, apply privacy, store in demo library
2. LEARN - Retrieval path (embed→index→search) + Training path (load→train→checkpoint)
3. EXECUTE - Observe→Policy→Ground→Act with Safety Gate and Evaluation

The Mermaid source is in the README for GitHub rendering, and this PNG is for documentation/offline use.