SCIP compiler-verified edges to raise depth≥2 trace confidence above heuristic tree-sitter

[100yenadmin]

Prior art: SDL-MCP's SCIP integration (scip-typescript, scip-go, rust-analyzer) — upgrades heuristic edges to "compiler-verified exact edges (confidence 0.95)."

Gap in GitNexus: call/import edges are tree-sitter heuristics. This is exactly why the usage discipline is "trust only d=1; d=2/d=3 are speculative — verify with rg." Heuristic edges cap confidence for multi-hop impact()/traces.

Proposed feature: optional SCIP ingestion for languages with indexers (scip-typescript, scip-go, scip-python, rust-analyzer). When a SCIP index is present, upgrade the matching edges from heuristic → compiler-verified, tag them with higher confidence, and let impact/query rank/return verified edges preferentially.

Value: directly addresses the documented depth≥2 reliability limitation — the single biggest correctness gap in current graph traces. Compiler-verified edges would make impact() trustworthy beyond d=1 for SCIP-covered languages.

Note: keep tree-sitter as the universal fallback (SCIP coverage is per-language and needs a build); SCIP is an enrichment, not a replacement.

[100yenadmin]

Source audit confirms: clear-cut real gap, highest value of the three — KEEP.

GitNexus is tree-sitter-only (heuristic edges, confidence caps ~0.9 → hence the "trust d=1" discipline); no scip/lsp/rust-analyzer anywhere in src. SDL ships a full provider-first SCIP + LSP pipeline (/indexing/providerFirst/lsp/mode, "SCIP index generation and ingest", same-run legacy fallback).

Clean fit: GitNexus already has the confidence-edge framework SCIP would slot into — per-edge confidence (0.5/0.8/0.9 in call-processor.ts), MIN_TRACE_CONFIDENCE/MIN_CONFIDENCE_LARGE gates, and a confidence-ranked resolver (resolve-references.ts). SCIP becomes a 0.95-confidence source feeding the existing resolver, tree-sitter staying as universal fallback. ~90% beneficial. CAVEAT: high implementation cost (per-language indexers + builds + provider-first/fallback plumbing) — a real project, not a quick win.