[smishra3]

What changed

The progress hook for the first-time weights download (src/subcellseg/weights.py) is now a small _Progress class that writes only when the shown value changes:

• in a terminal: the line is redrawn in place once per whole percent (101 redraws);
• when stderr is not a terminal (CI, nohup, SLURM logs, notebooks): one line per 10% (11 lines);
• unknown size: a terminal redraws once per MB; a non-terminal writes nothing until the download ends.

Download, SHA-256 check and caching are unchanged.

Fixes #5.

Why

urlretrieve calls the hook for every 8 KB block, about 8,000 times for the 64 MB weights, and each call rewrote the line. Any captured log got ~150 KB of 0 / 64 MB (0%) text.

How it was tested

• pytest: 13 passed on Linux (CPU), including 2 new cases checking the output for both the terminal and non-terminal paths.
• Real download from the public release with an empty SUBCELLSEG_CACHE_DIR:

Output	Before	After
stderr to a file	~150 KB, ~8,000 redraws	407 bytes, 11 lines
under a terminal (script)	~8,000 redraws	101 redraws

The downloaded file's SHA-256 is unchanged (858689a3…). Segmentation code is untouched, so no bench_predict.sh run is needed.

Checklist

• Tests pass locally (pytest)
• Docs updated if behaviour or commands changed (no command changes)
• CHANGELOG.md has an entry under Unreleased (if a user would notice)
• No weights, outputs or notebook outputs committed
• AI-assisted code is reviewed line by line and the commit carries an Assisted-by: trailer (written with Claude Code; for the maintainer to review)

Assisted-by: claude-code (claude-opus-5-5)

🤖 Generated with Claude Code