docs: document incremental builds, dashboard, entity resolution

The README described a tool that no longer matches what ships. Adds the
features built since, and a "What Changed Recently" section summarising
the retrieval work.

New sections:
  - Incremental, versioned builds (--rebuild / --prune, resume semantics,
    corpus_version on /health)
  - Built-in dashboard at / (books, graph explorer, retrieval tester)
  - Entity resolution, including --llm and the hand-edits-win rule
  - agent.yaml domain configuration, with a callout that
    extraction.predicates is a closed vocabulary that DROPS unmatched
    facts — the one setting that silently breaks a new-domain corpus

Updated the runtime half of the architecture diagram, which omitted query
rewriting, traversal and citations.

The changelog section is collapsed by default and states what was wrong
rather than only what is new, since several items are the kind of silent
failure worth recognising elsewhere: a graph scan that stopped after 30
matches and made later documents unreachable, chunk auto-tuning that
produced ~115,000-character chunks, and a vector store that reported zero
documents after reopening.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>