feat(knowledge): expand category taxonomy + single source of truth (P3 stage 1)

[mkreyman]

What

P3 stage 1 of the second-brain taxonomy work: expand the category set and centralize it as a single source of truth.

The corpus only had 5 categories (pattern/convention/decision/finding/reference). This adds the richer taxonomy you approved and — just as importantly — kills the 6-way duplication of the category list that was silently drifting.

Taxonomy

New module Loopctl.Knowledge.Categories is canonical:

• active/0 — pattern, decision, finding, reference, playbook, insight, entity, idea, quote, question (the 4 kept + 6 new). What new content is classified as.
• retired/0 — convention. Still DB-valid (the category column is a plain string and ~13% of the corpus is convention; dropping the enum value before the reclassification backfill would make those rows fail to load), but not offered to new content.
• definition/1 + prompt_fragment/0 — per-category definitions injected into the LLM extraction prompts so the model classifies into the full taxonomy correctly (e.g. playbook = the steps vs pattern = the shape).

Wired through (drift eliminated)

Before (own copy of the list)	After
Article.@category_values	Categories.all() (Ecto.Enum can't drift)
ClaudeContentExtractor prompt + @valid_categories	prompt rebuilt from Categories at compile time; gate → Categories.all_strings()
LlmExtractor prompt + @valid_categories	same
ReviewKnowledgeWorker.@valid_categories	Categories.all()
ContentIngestionWorker.@valid_categories	Categories.all()

scale_seed.ex is intentionally left alone — it's synthetic scale-test data whose category round-robin feeds index-selectivity gates; changing its list length is orthogonal to the product taxonomy and would risk scale-gate flakiness.

No DB migration: category is a :string column, so new enum values need no schema change.

Tests

• New categories_test.exs: composition (all = active ++ retired, disjoint), the expanded active set, string helpers, definitions cover every active category, prompt fragment includes all active / excludes retired, and an enum-vs-canonical drift guard (Ecto.Enum.values(Article, :category) == Categories.all()).
• article_test.exs: changeset accepts every new category.
• knowledge_test.exs + knowledge_stats_controller_test.exs: the dense by_category stats maps grew from 5→11 keys (correct — stats derives the dense map from the enum). Updated the expectations to derive from Categories rather than re-hardcode, so a future taxonomy change can't silently break them again.

Full local gate green: format, credo --strict, dialyzer (0 errors), 2987 tests / 0 failures.

Next P3 stages (separate PRs)

1. CategoryClassifier behaviour + config-DI (Claude-backed) + Mox.
2. Resumable, cost-ceilinged, dry-run-first backfill worker over the 77k corpus: write-on-confident-change, always reclassify convention; emits a proposed-distribution report before committing.
3. Cross-repo: synth.py / publish.py category enum (claude-config) + user CLAUDE.md category list.
4. After the backfill drains convention to 0, a follow-up drops it from active/retired.