feat: retrieval ladder v0.5.0 (SQLite FTS5 + ripgrep)

[RichardHightower]

Replaces the JSON knowledge/.index/ inverted index (committed by accident in 0.4.x) with PKC's disposable retrieval ladder. Git + Markdown stays source of truth.

index present  →  use it
else rg on PATH  →  prefilter
else             →  pure Python scan

What landed

• SQLite/FTS5 incremental index at knowledge/.dekc/index.sqlite (gitignored). Search and pack self-heal via mtime+size. build remains an alias for refresh --force so CI and existing skills keep working.
• ripgrep prefilter (DEKC_RG_PATH / PKC_RG_PATH / OKF_RG_PATH). Optional. Never installed from a hook.
• Score identity: Python scorer (title×10 / desc×5 / tags×4 / min(body,8)) ranks candidates. LIKE on the stored haystack is the default; --engine fts is the FTS5 MATCH opt-in (not score-identical).
• Pack identity: undirected lineage (typed flow + SQL) for every visited node; extra non-lineage links[] still attach to the focus only. Pack always parses the current file for outbound SQL — rg -lF <path> cannot see a file that does not mention itself.
• SQL edges store origin so incremental delete does not leak table-to-table edges.
• Host manifests lockstep at 0.5.0 (.opencode/plugin/dekc.json and public/data/catalog.json were still 0.4.2).
• Removes the committed sample-knowledge/.index/ JSON.

Design: docs/designs/retrieval-ladder.md.

Tests

• tests/test_retrieval_ladder.py — rg identity, sqlite refresh/noop/touch/delete/drop, no JSON written, doctor path, sample revenue hits
• existing tests/test_dekc.py (incl. version lockstep + plugin.json)
• CI now globs scripts/dekc_*.py and asserts search engine plus index fts5/nodes

Leftovers (accepted)

• dekc_pack.py --mermaid still calls build_graph (full scan)
• Do not ignore the whole .dekc/ directory — .dekc/config.example.yml is committed