Section self-containedness: chunk-quality signal for get-section consumers

Why

matlatl serve exposes get-section — agents consume individual sections as context chunks. A section whose prose depends on unstated context ("as configured above", "using the same approach", a table continued from the previous section) is a bad chunk: served alone, it misleads.

This is the one heuristic in the agent-first set that's genuinely novel (chunk quality for retrieval is well-studied in RAG, but not as a deterministic lint over authored markdown).

What to build (lexical, deterministic — no NLP)

Score each section on cheap self-containedness proxies:

1. Anaphoric openers — section body begins with a referring phrase from a small in-source set ("as above", "as mentioned", "this approach", "the same", "these", "it" as first token…), mirroring scent.go's phrase-set pattern.
2. Dangling definitions — the section uses backticked identifiers/acronyms that are defined (first bold/heading mention) only in sibling sections of the same doc. (Scope carefully — start with openers only if this is too noisy.)
3. Surface as a context-dependent-section Info finding (never gating) + optionally per-section data in graph.json.

Sequencing

Lowest priority of the agent-first additions: highest false-positive risk, and its value depends on get-section actually being used by agents — let the P13 eval establish that first. Dogfood extensively; be prepared to keep it behind config if noisy.

🤖 Generated with Claude Code

Start by reading scent.go's phrase-set pattern and tracing the get-section consumer path. Implement the lowest-risk opener heuristic first, then dogfood it on markdown sections; done means context-dependent-section Info findings are non-gating and false positives are assessed before adding dangling-definition or graph.json data support.