ci: compatibility sweep against public OME-Zarr datasets

[MPanknin]

A compatibility sweep: how much of the public OME-Zarr data opens in Kiln, and why the rest doesn't.

• What it checks: the viewer's own code (metadata parsing, validation, pyramid fallback, level limit), then fetches and decodes one chunk per dataset. It also checks CORS and, for HCS plates, whether the first field opens on its own. Nothing is rendered.
• Datasets: up to 15 per list, sampled evenly, from the OME-NGFF Challenge 2024 lists, the IDR OME-NGFF samples, the OME-Zarr Open SciVis collection and the Kiln gallery. 336 datasets, about 3.5 minutes.
• Runs: from the Actions tab (Compatibility sweep → Run workflow, with an optional sample size) and on pushes that change the sweep.
• Results: REPORT.md and results.csv on the compat-sweep branch, one commit per run, plus the job summary.
• Code: scripts/compat-sweep/ only. The Zarr provider's Vite worker import resolves to a stub under bun, so no library code changes.

First run: 159 of 336 open (47%). HCS plates are 44% of the sample, and 75% of the volumes that open store one z-slice per chunk.

Checked locally against test stores covering each outcome: v0.4/v0.5, sharded, uint32, plate, time series, 6 channels, bioformats2raw, labels, no CORS, 404, URL without .zarr.

---

Generated by Claude Code