fix: align policy identity normalization with grammar contract

[rlippmann]

What changed

• Narrowed policy identity normalization to:
  • Unicode NFKC normalization
  • apostrophe glyph normalization
  • case-insensitive matching
  • whitespace collapse and trimming
• Removed leading article stripping for a, an, and the.
• Removed the English-specific dont → don't rewrite.
• Promoted the existing spec-target expectations into active conformance coverage.
• Updated engine, property, step, and state JSON fixtures to preserve distinct policy identities such as:
  • docker and the docker
  • dont and don't

Why

• The runtime applied language-specific interpretation that was not part of the documented policy identity contract.
• Article removal and contraction correction belonged to older acquisition-oriented behavior, not deterministic identity normalization in the authority layer.
• Aligning runtime behavior with the existing grammar specification makes policy identity predictable and keeps the core normalization rules language-neutral.

Checklist

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