4/ The JSON tree rendered every node: 200,002 rows and 1.4M DOM nodes for a 20k-item file. Virtualized, it's ~80 rows. The lesson: my benchmarks only tested easy inputs. Profile the ugly ones.

3/ Two sneaky O(n²) loops: checking item order in JSON arrays (20k items: 490 → 96 ms), and word overlap on long lines. A 100k-word minified line went from 102 s to 0.23 s.

2/ My PDF text extractor rebuilt every font, re-parsing its ToUnicode map, on every single page. Caching fonts per document made extraction 4× faster, with byte-identical output across my PDF test corpus.

1/ Tabs now remember their result. Switching back checks the files' fingerprints (a few ms) and only compares again if something changed on disk. Result on screen in under 50 ms instead of seconds.

The black-and-white treatment suits the photos, and the two categories are immediately clear. The logo is hard to read at that size, though. I’d make it larger or move “pets, people, events, prints” into normal page text.

Much better default. The group labels and indentation make the long menu easier to scan, while the ungrouped select still looks compact. This is the first version I wouldn’t immediately restyle.