feat(research-library): library-acquisition skill + consolidated PDF extractor

[Fede654]

Research Library — citable primary-source corpus

Adds a Research Library capability to HMK: a filesystem corpus of full primary
sources (books, papers), citable as (Author, Year, ch. X, p. Y), kept distinct
from the memory store (agent-memory/library.db). The two were being conflated
under the word "library"; this PR separates the concepts and bridges them.

What's included

• scripts/library_extract.py — a single PyMuPDF extractor with two modes:
  • corpus: structured ingestion (copies the PDF to raw/, extracts + cleans the
    full text, writes meta.json, updates topic + master indexes).
  • dump: flat full-text extraction (every page) to stdout/file with a
    pages=N chars=M report — for one-off translation/TTS of a PDF not in the corpus.
  • Replaces two per-agent scripts (structured vs flat) that had drifted. Paths are
    workspace-relative ($LIBRARY_ROOT / $HMK_WORKSPACE_ROOT/library).
• templates/skills/research/library-acquisition/ — the acquisition procedure,
  generalized from a per-agent local skill. Its closing step registers a catalog
  pointer in library.db (shelf evidence) pointing at each book's meta.json, so
  corpus books are discoverable via librarian/hybrid-pack and expand jumps to the
  on-disk chapter. This is the memory↔corpus bridge.
• librarian skill — cross-links library-acquisition and documents the corpus as
  a third retrieval tier, disambiguating the two "libraries".

Model

library.db holds the index (pointers/summaries); the Research Library holds the
full text. Acquisition feeds discoverability; the wiki map remains the orientation layer.

Origin: generalized from an agent-local Research Library + acquisition procedure.

🤖 Generated with Claude Code