Add provenance-preserving project structure graph adapters

[AjnasNB]

Goal

Add deterministic, provenance-preserving project-structure adapters that turn code, schemas, and selected documents into compact queryable relations without copying another project's implementation or turning derived graph output into authority.

Acceptance criteria

• Start with language-aware AST/import/symbol extraction for a narrow reviewed language set; every node and edge cites an exact file identity and source span.
• Mark relations as extracted or inferred. Inference requires an explicit adapter/model identity, confidence, source set, and reproducible prompt/config fingerprint.
• Store source observations as normal Qarinah events; graph/community/index files remain disposable derived projections.
• Incremental rebuilds detect additions, changes, renames, and deletions by content identity and cannot leave stale authoritative-looking edges.
• Hybrid retrieval can combine project structure, temporal decisions, crawler evidence, and Maqam receipts under one token budget while preserving conflicts and supersession.
• Secrets, ignored files, binaries, generated directories, hidden reasoning, and paths outside the trusted real workspace root are excluded by default.
• Offline fixtures cover cycles, duplicate symbols, aliases, malformed files, huge files, symlinks/junctions, rename history, poisoned indexes, and deterministic clean rebuilds on Node 22/24/26.
• Codex and Claude skills expose scoped query workflows; they do not auto-inject the whole graph into every prompt.

Non-goals

No universal semantic understanding, no replacement for compilers or language servers, no unbounded repository ingestion, and no claim that inferred relations are source truth.

[AjnasNB]

PR #25 closes one foundation gap by materializing every relation target as a typed graph node, so event graph edges no longer dangle. The existing bounded scanner already covers paths, hashes, JS/TS imports, Markdown links, spans, additions, changes, renames, and deletions.

Keep this issue open for the deeper adapter work: versioned JS/TS symbol extraction first, then calls/inheritance/implementation edges, queryable file/symbol nodes, incremental content-hash rebuilds, and impact retrieval. The lightweight zero-dependency file graph remains the fallback.