Enrichment pass with a local language model

Section 10 of `docs/extractium-spec.md` describes a deferred enrichment step: a small local instruct model, GPU-gated, delta-only, and parent-only, populating the reserved nullable fields (`summary`, `tags`, `keywords`, `enriched_at`, `enrich_ver`) through a `.kb_cache/enrichment/` layer, with no API calls ever. The fields are reserved in the container format and every adapter writes them when present and skips them when null, so the format needs no change.

Keyword extraction without a language model (YAKE plus embedding-similarity keywords) may ship earlier as a separate step, and would be a smaller first phase. Either way, plan it as a phase in `docs/implementation-plan.md` before building.

Read Section 10 of docs/extractium-spec.md and inspect docs/implementation-plan.md first. Document the enrichment phase, including the GPU gate, delta-only and parent-only scope, reserved nullable fields, .kb_cache/enrichment/ layer, and no-API constraint; clarify whether YAKE and embedding-similarity keywords are a separate earlier phase. Done means the implementation plan records an agreed sequence and boundaries before coding begins.

Written by the indexing model from the issue text.