Local AI (llama.cpp) layer: embeddings matching, issue normalization, content-quality flags, nightly digest

[mgifford]

Umbrella issue — implement as separate bounded sessions in this order:

1. Embedding-based issue matching (S/M) — llama-server /embedding endpoint (nomic-embed or bge-m3, CPU is fine). Embed open-issue title+body (cached, re-embed only on change) + our pattern summaries; cosine similarity replaces the keyword scorer in sync-known-issues.js. Upgrades Known-issues reconciliation v2: body matching, issue-map upkeep, staleness nudges #38.
2. Issue-queue normalization (M) — GBNF grammar-constrained JSON extraction per known issue: states a WCAG SC? has repro steps? names selector/template? Output 'issue quality cards' driving targeted nudge drafts (missing SC → suggest wcagXXX tag, etc.).
3. Content-quality flags (M) — a detection class rules cannot do: useless alt text ('image.png'), ambiguous link text in context, headings that don't describe sections. Nightly pass over crawl DOM; everything flagged as 'needs human review', never auto-filed.
4. Nightly digest (S) — 5-sentence human summary of each night's diff (new patterns + attribution, verified fixes, trends) from structured input only.
5. Draft prose assist (S) — impact statements for issue drafts and OpenACR remarks from pattern data into templates; labeled as drafted.

Hard guardrails (non-negotiable): no conformance determinations, no auto-merging of patterns, no auto-posting anywhere; every AI task needs a calibration accuracy figure (extends #27) before its output is trusted; batch nightly + cache embeddings + smallest model that passes calibration.

Runs on the nightly server's llama.cpp (llama-server: OpenAI-compatible API, /embedding, GBNF grammars). Aligns with the local-first Tier C routing in the owner's governance docs.

---

Drafted with AI assistance (Claude Code); no external code copied.