refactor: simplify LiteLLM DirectiveDrafter example boundary

[rlippmann]

Summary

Refactor the LiteLLM prompt-construction DirectiveDrafter example to consume the current DirectiveDrafter public boundary instead of duplicating drafter behavior in the example.

This PR updates the example to:

• use DirectiveDrafter.draft_directive(...)
• provide only a LiteLLM fallback callback that returns raw candidate text
• consume DraftResult variants explicitly
• read compiled state directly from engine.premise and engine.policies
• remove example-owned directive grammar matching, resume/checkpoint behavior, and other runtime-style workflow logic

Why

The example had drifted away from the intended consumer boundary and was still re-implementing pieces of drafter behavior locally. That made the example misleading as documentation and also masked live behavior issues during end-to-end testing.

This change keeps the example focused on the provider/drafter/compiler handoff and updates the tests to verify the public API seam instead of internal helper behavior.

Example metadata

• Enforcement point: host-side prompt construction from authoritative compiler state
• Domain: generic assistant request shaping
• Technology/runtime: Python, LiteLLM, DirectiveDrafter, Ollama-compatible provider flow
• Example type: cookbook
• Does this use directive drafter? should normally be no: yes
• Does this derive state from model output? must be no for primary examples: no

Checklist

• Repository scaffolding / governance change
• One primary enforcement point
• Explicit authoritative state
• No model-derived state mutation
• Adversarial stub or equivalent test
• Observable runtime behavior change
• Domain vocabulary is natural
• Framework is secondary to the enforcement point
• Docs updated
• Tests/smoke checks added where practical

Notes

Focused validation run for the changed files:

• uv run --no-sync pytest python/tests/test_litellm_with_directive_drafter.py
• uv run --no-sync ruff check python/examples/prompt_construction/litellm/with_directive_drafter.py python/tests/test_litellm_with_directive_drafter.py
• uv run --no-sync mypy python/examples/prompt_construction/litellm/with_directive_drafter.py python/tests/test_litellm_with_directive_drafter.py

Live testing against local Ollama on Monday, August 10, 2026 exposed that the example still over-applies drafted directives on some ordinary conversational turns. This PR keeps the boundary cleanup and test refactor in place, but the live test results indicate there is still follow-up behavior work to do in the example flow.