test: harden demo oracles and clarify results docs

[rlippmann]

What changed

• Hardened demo oracle/checker logic against wording-variant false positives in scored demos:
  • Demo 02 prohibited-content negation handling
  • Demo 03 stale-premise negation handling
  • Demo 04 tool tag normalization robustness
  • Demo 05/07 premise-tag normalization robustness
• Expanded oracle-focused tests (including Hypothesis/property coverage) and baseline/comparison parity checks across demo paths.
• Updated documentation to make demo evidence easier to discover while reducing duplication:
  • Added a concise top-level "Does it work?" summary in README.md
  • Added canonical demo-results reference page: docs/demos-results.md
  • Linked README.md and demos/README.md to canonical results
  • Replaced weak Demo 05 example with a real long-context drift example snippet
  • Removed redundant/stale evidence summaries from README.md

Why

• Live runs exposed repeated false failures caused by brittle wording-sensitive oracles rather than engine behavior.
• The oracle hardening and property tests reduce regressions from harmless text variants while preserving strict failure behavior for genuinely unsafe/incorrect outputs.
• Documentation updates provide a single source of truth for results and improve discoverability without duplicating matrices across files.

Checklist

• pre-commit run (uv run pre-commit run --all-files)
• tests pass (uv run pytest)