Rewrite README for outsiders

[muneebshahid]

What

Closes TIL-9 — the README is reframed for a stranger evaluating the project, per the ticket's acceptance criteria:

• Positioning: leads with "runtime, not a framework" — caller-constructed dependencies, no plugins/global config/hidden defaults.
• Install + quickstart: pip install tile-runtime (dist name vs tile import name stated explicitly, since bare tile is taken on PyPI), then the embedding quickstart — the exact snippet proven in TIL-8's clean-venv packaging run.
• Caveats up front: 0.x stability disclaimer, OpenAI-only, Python 3.13+.
• Typed results section (from TIL-18): result=Schema example with match on Completed/Failed, the per-prompt complete/fail mechanics, reserved names, and the schema design guidance — demand judgment not transcripts, bulk data on ToolResult.details, summary: str for guaranteed prose.
• Status/outcome contract: the invariant (outcome non-None ⟺ status == "completed"), the six-row terminal-state table, failure recovery (history at last stable state, healing, session re-promptable), and the no-retries position (SDK max_retries + re-prompting).
• Prompt-caching note (from TIL-20): reuse one result schema per session; typed/plain flips re-read the session history at full price.
• New cwd surface from TIL-23 documented in the quickstart: required cwd, unbound BUILTIN_TOOLS, declare-cwd-to-receive-it custom tools.
• CI badge retained; Security Posture kept verbatim; architecture tree updated (agent.py, prompt.py, result.py).

Verification

Both code snippets were executed against the live OpenAI API (gpt-5.4) from the TIL-8 clean venv using the built 0.1.0 wheel, under env -i with only OPENAI_API_KEY: the quickstart completed with correct tool-driven output, and the typed-results snippet returned a validated WeatherReport through the match arms as written.

🤖 Generated with Claude Code