Refresh OCR model lineup: local-first routing, drop dead-weight engines

[r-uben]

Problem

The OCR engine lineup no longer reflects measured quality/cost on the owner's hardware and born-digital academic workload. Several engines are mis-ranked or dead weight:

• DeepSeek-OCR scores ~0.085 (socOCRbench, near dead last) yet sits in AUTO_ENGINE_ORDER and _LOCAL_ENGINE_ORDER.
• Mistral OCR (~0.45) is ranked as the premium rung above Gemini, but it is worse AND ~5x pricier than Gemini — strictly dominated.
• Figure description is hardwired to the Gemini API with no local-first path.
• judge / math-recovery defaults lean on slow/timeout-prone local models instead of the practical winner qwen3.5:cloud.

Policy

Best cheap combination, local-first → Ollama Cloud → Gemini/Mistral edge-case only. Native PyMuPDF stays the real default for born-digital text and table values.

Measured winners (owner's 64GB Mac, real workload)

• qwen3.5:cloud (Ollama Cloud, no extra key): ~0.57, ~49s/page, only engine to clear all 3 hard page types → practical workhorse VLM.
• Gemini 3.x: best quality (0.60-0.64) but paid + occasional empty pages → edge escalation only.
• DeepSeek-OCR / Mistral: drop from auto ladders.

Scope (this issue)

1. New AUTO_ENGINE_ORDER: QWEN → GEMINI → MARKER → GLM → NOUGAT (drop DeepSeek + Mistral from auto).
2. Demote Mistral to manual-only fallback; drop DeepSeek from _LOCAL_ENGINE_ORDER.
3. Flip judge_model / math_model defaults to qwen3.5:cloud.
4. Add an LLM-friendly MODELS.md documenting the routing policy per sub-task (text/formula, figure description, table layout).

Follow-ups (separate)

• Local-first figure description in gemini_api.py.
• Split provider identity by engine+backend+model (QWEN local vs cloud) + fix QwenEngine.is_available() cloud gating.

Design second opinion: Codex (gpt-5.5) — agreed with the lineup above.