fix: markdown→HTML rendering in engine + harden thinking fallback

- Engine now converts content_md to HTML at render time (was dumping raw markdown,
  causing articles to show literal #/**/- symbols and collapse into wall of text)
- /api/publish accepts 'content' key and converts markdown→HTML for API consumers
- Added md Jinja filter + md_to_html helper (markdown lib, extra+sane_lists)
- orchestrator: log warning when falling back to 'thinking' field (CoT, not prose)
- content_pipeline now generates formatted articles via LLM instead of raw scraped HTML

core/orchestrator.py

@@ -205,9 +205,14 @@ def llm_chat(prompt: str, model: str = "qwen3.5:4b-mlx", host: str = OLLAMA_MACB
                     result = r.json()
                     if "message" in result:
                         content = result["message"].get("content", "")
-                        # ornith puts output in 'thinking' when content is empty
+                        # ornith puts output in 'thinking' when content is empty.
+                        # WARNING: 'thinking' is chain-of-thought reasoning, NOT article text.
+                        # Only fall back to it for JSON/short tasks, never long-form prose.
                         if not content:
                             content = result["message"].get("thinking", "")
+                            if content:
+                                log.warning(f"LLM {model} returned empty content — fell back to 'thinking' field ({len(content)} chars). "
+                                            f"Verify this is real output, not chain-of-thought.")
                         if content:
                             return content
                     if "error" in result: