Audit class-reference ground truth exposed by exact Bifrost ranges

[DavidBakerEffendi]

Context

The exact-range Bifrost run used for PR #62 reports three extra class-reference locations that may be legitimate source references rather than analyzer false positives:

• csharp-class-reference: src/Consumer.cs:8:31, the Service token in new Service(repository).
• python-barrel-class-construction: consumer.py:6:13, the Child type annotation in client: Child = Child().
• python-multilevel-barrel-class-construction: consumer.py:8:17, the Grandchild type annotation in grandchild: Grandchild = Grandchild().

The Java baseline already treats constructor tokens as class usages in its class-reference case, while the C# and Python cases currently narrow their authored expectations to one syntactic role. These cases need semantic adjudication before changing either the corpus or Bifrost.

Required review

• Compare standard LSP reference behavior for each declaration.
• Decide whether class declaration scans cover all semantic class references or may be narrowed to construction-only/type-only roles.
• Keep declaration-to-usage and constructor-specific cases internally consistent across languages.
• Do not add allowed extras solely to improve the score.

Acceptance criteria

• Each of the three locations is explicitly accepted or rejected with rationale.
• Reviewed case notes and expected locations reflect the decision.
• cargo run -- validate benchmarks/cases passes.
• A full Bifrost report records the resulting exact status.