docs: refresh demo evaluation notes and provider guidance

[rlippmann]

What changed:

• Added local Ollama context-size sweep results for the four-path demo comparison in docs/demos-results.md.
• Added an exploratory Demo 05 long-transcript frontier note (separate from the historical full matrix).
• Clarified direct Anthropic vs LiteLLM gateway OpenAI-compatible model naming in demos/README.md and docs/demos-results.md.
• Updated stale demo wording that referred to three-path outcomes where four paths are now reported.

Why:

• Keep demo docs aligned with current evaluation paths (baseline, reinjected-state, compiler, compiler+compact).
• Prevent confusion around Anthropic model naming across direct endpoint vs gateway usage.
• Preserve historical matrix integrity while documenting new exploratory evidence separately.