Build deterministic Tree-sitter DiffGraph baseline with evidence

[nia-sg-bot]

Why

The canonical DiffGraph topology must be reproducible and useful without vendor credits or an LLM call.

Direction

• Git pre/post blobs provide source material.
• Tree-sitter provides the deterministic baseline.
• LSP may enrich resolution when available.
• AI may explain the frozen artifact but cannot add/remove nodes or relationships.

Acceptance criteria

• Produce schema-valid symbols and relationships directly from exact pre/post blobs.
• Start with Python, with explicit capability gaps for unsupported languages.
• Stable IDs/order for identical inputs.
• Evidence and analyzer/parser/query provenance on every structural finding.
• Parse failures become scoped warnings, never silent empty snapshots.
• Explicit unresolved/external entity semantics.
• Golden fixtures for add/modify/delete/rename/import/call/containment and parser failure.
• Tests prove optional prose enrichment cannot mutate topology.

Related

Architecture issue #8 and the implementation work in PR #13 must be reconciled against this contract.