P12 — Semantic frontier (opt-in): embedding-based 'similar but unlinked' gaps

What ships

The strongest 'these are related' detector: embedding-based semantic similarity to surface topically-similar-but-unlinked doc pairs — the gaps that topology alone (P8) can't see.

Why / constraint

Most powerful gap signal, but non-deterministic and model-dependent. It MUST stay behind a flag and out of the default golden/byte-stable output path, exactly the way --check-external keeps network results out of deterministic output today. Document it as the explicit boundary of the deterministic core. Lowest priority — P8's topology-only methods deliver most of the value deterministically first.

Where

New opt-in analyzer + an opt-in dependency (local embedding model or lexical TF-IDF fallback over headings as a deterministic middle ground). Flag-gated; never in default output.

Acceptance

• Off by default; only runs behind an explicit flag
• Results kept out of the deterministic default output (like --check-external)
• Determinism boundary documented in an ADR
• Consider a deterministic lexical (TF-IDF over headings) tier before pulling in embeddings
• Theory: Salton vector space model; Deerwester et al. 1990 (LSI); modern embeddings

Part of the navigability roadmap derived from docs/research/information-organization-theory.md (plain-language: docs/research/information-organization-explained.md).

Start with docs/research/information-organization-theory.md and docs/research/information-organization-explained.md, then inspect the existing --check-external flag and P8's topology-only methods. Define an opt-in analyzer, keep its results out of default golden/byte-stable output, and document that determinism boundary in an ADR. Consider TF-IDF over headings before adding embeddings.