ETL-style parallel markdown backfill with real-time progress

- ProcessPoolExecutor: 1 process per CPU core (bypasses GIL)
- ThreadPoolExecutor: 6 threads per process for I/O throughput
- Thread-local SQLite connections with WAL mode
- Manager().Value() for cross-process progress counter
- Real-time tqdm updates polling shared counter every 50ms
- ~19 pages/sec on 4-core system (4887 pages in 2 min)

html2md.py

@@ -426,8 +426,12 @@ class SmartMarkdownConverter:
                     src = self._resolve_url(child.get('src', ''))
                     alt = child.get('alt', '')
                     if src:
-                        # Images should be block-level, not inline
-                        parts.append(f"\n\n![{alt}]({src})\n\n")
+                        # Emoji images (alt like :smile:) stay inline
+                        if alt.startswith(':') and alt.endswith(':'):
+                            parts.append(f"![{alt}]({src})")
+                        else:
+                            # Regular images are block-level
+                            parts.append(f"\n\n![{alt}]({src})\n\n")
 
                 elif tag == 'br':
                     parts.append('\n')