[kvlonge]

Why

The preview engine still resolves each model's declarations in Python. Working out which enums,
constants and macros a model can see, and finding its @enum/@const references, is lookup and
text-scanning work that the native scope index and SQL scanner can do without changing results.

Changes

• Register a preview-tier native stage, model_loop, for the model compile loop.
• Classify each indexed consumer's declaration context natively from the scope lookup (global,
  private, local, inherited and relationship-granted declarations) and build the same Python
  context objects. Unindexed paths, and path targets the scope query reads as qualified
  identities, keep the Python resolver and its errors.
• Scan the @enum/@const references of every model in one native batch using Python's quote,
  comment and dollar-quote rules. Values are still rendered in Python; malformed references,
  unclosed text, Unicode word boundaries, unknown names and render errors fall back to the
  Python expansion, which reports the exact error.
• Native context classification runs inside the compiler panic boundary.
• Native model config mismatches now raise the shared NativeStageMismatchError.
• Stage captures can omit declared dataclass fields, keeping the native context handle out of
  the compiler differential.

Verification

• make compiler-differential and make compiler-differential-shipped identical; compiler
  differential e2e tests pass.
• Seeded parity tests for declaration contexts and 3,000 generated SQL strings; CLI parity tests
  for folder-private references and a model folder containing :.
• make check-ci, make test, Rust clippy, fmt and workspace tests.