Phase 4: --bakeoff CLI, backfill/docs polish (#9), AO3 vocab seed (#8)

[elfensky]

Phase 4 of the desloppify code-health roadmap. Closes the last two open roadmap issues. Full suite green (42 core tests).

#9 (final items) — 68a6619

• scourgify classify --bakeoff exposes the engine comparison that was locked in the wizard (display-only, writes nothing). Engine metadata (ENGINE_ENV + usable_engines()) now lives in classify as the single source of truth; wizard.ENGINE_KEYS aliases it so the two can't disagree.
• Backfill preview shows each book's title, not just a raw id.
• README engine list now includes mistral (already present in PRICING/ENGINE_ENV/classify).

#8 — seed the controlled vocab from AO3 (0fd851c)

The curated vocab was ~120 terms while wrangle ships the full ~93k-fold AO3 taxonomy, so the classifier under-tagged out of the box. The highest-use AO3 freeforms are the generic reusable tropes (Fluff, Angst, Slow Burn, Hurt/Comfort…) — a use-count threshold cleanly separates them from fandom-specific noise (which never accumulates many uses).

• New maintainer tool build_classify_seed.py reads data/ao3_vocab.csv (name,uses, from the OTW dump via build_ao3_layer.py), gates at ≥20k uses, drops fandom-specific (...) and format/meta tags, and writes the regenerable defaults/classify_vocab_ao3.txt (~120 terms).
• load_vocab() now merges curated core ∪ AO3 seed (dedup-on-merge, curated spelling wins), still fully overridable (-term trims a seeded term too).
• The merge-dedup test surfaced and fixed a pre-existing Found Family duplicate in the curated file.
• Verified the new seed ships in the wheel; data/ does not leak.

🤖 Generated with Claude Code