Measured desktop baseline — 2026-10-14

Hardware: AMD Ryzen 9 8940HX, approximately 23 GiB usable RAM, Windows 11 build 26300, Python 3.11.9. GPU is not used for these operations. Synthetic files are generated in a temporary directory by scripts/benchmark.py and removed afterward.

Measurement	Result
Main window construction through first processed paint	86.97 ms
Imported JPEG files	300
Total source bytes	403,689,900 (~385 MiB)
Import time (metadata + verification + SQLite batches)	1.070 s
Import throughput	280.5 images/s
1920×1080 decode median, 30 samples	22.24 ms
Decode P95 (29th sorted sample)	25.03 ms
Metadata records inserted + paged through	100,300
Metadata insertion + scrolling	1.530 s
Metadata pages retained	8
Metadata records retained	972 (maximum 1,024)
Process working set before metadata exercise	76.13 MiB
Process working set after metadata exercise	81.03 MiB

Window construction excludes interpreter/Qt imports. Decode timing measures worker work, not full click-to-display latency or a cold disk. The 300 JPEGs are copies of one synthetic noisy image, so cache effects matter. The additional 100,000 records are explicitly metadata-only fixtures; they are not 100,000 real decoded images. Working set is measured with Windows GetProcessMemoryInfo, not private bytes.

Imports and exports run outside the UI thread. The list is virtualized with an eight-page metadata cache; decoded preview cache has a 96 MiB budget and serial decoding. There is no claim of a measured 10 GB real dataset, cold startup, clean-VM performance, or cross-machine sync latency. Those require representative user data and the target network. No comparative ranking is claimed.

Reproduce from the repository with python scripts/benchmark.py after installing the package. The script reports hardware/runtime and raw metrics as JSON.