docs(tutorial): restore the napari screenshot cells and result images

[xuefei-wang]

Problem

A user reports that the tutorial "stripped out the important steps to run/visualize the results," comparing against 616d4b5.

They're right. The v0.1.0 monorepo merge (4a2d7ba) replaced both napari screenshot cells and both embedded result images with prose:

The image and segmentation layers appear directly in the interactive Napari viewer. Static documentation builds do not execute or embed GUI screenshots.

So the tutorial's two visual payoffs — the multiplexed image with the CellSAM segmentation overlaid, and the per-cell-type label layers — disappeared from the rendered page. Both "Visualizing results" sections now end with no output at all.

Why the premise was wrong

napari does render headlessly (Xvfb + llvmpipe), and that is how the currently-deployed site was built. https://vanvalenlab.github.io/deepcell-types/site/tutorial.html still serves both images today:

• _static/_generated/napari_img_and_segmentation.png
• _static/_generated/napari_celltype_layers.png

The repo source and the published build had silently diverged. This restores the source to match the build that actually works.

Change

Restores the two hide-cell nim.screenshot(...) cells and the two <img> embeds. Docs-only; no code touched.

Verification

• jupytext parses the tutorial: 45 cells / 24 code cells, 2 screenshot cells, 2 image embeds
• pytest: 470 passed, 1 skipped
• Resulting docs/site/tutorial.md is byte-identical to the version that produced the live site, apart from one unrelated abstention-wording sentence left untouched

Not addressed here

The same report also says the model won't run. Everything testable offline is healthy — pip install git+...@master resolves (incl. deepcell_auth from git), the deepcell-auth manifest maps 2026-06-15 to models/deepcell-types_2026-06-15_resmlp.pt / b819a7e0…, and predict() runs fine against that checkpoint on this branch. The one link that can't be checked without a token is whether the bucket actually serves an asset with that hash. Tracking separately.

Separately: gh-pages was last built 2026-07-12, which predates the deepcell-auth migration (#47, 2026-07-30) — the published site should be rebuilt from master once this lands.

[xuefei-wang]

Superseded by #55 — closing as redundant.

#55 contains this PR's changes verbatim: the same two hide-cell nim.screenshot(...) cells and the same two <img> embeds for napari_img_and_segmentation.png and napari_celltype_layers.png. Verified by diffing this branch against #55's head — the only differences are #55's additional empty-mask guard; nothing from this PR appears as a removal.

Both changes were validated together on an integration branch merging the two (clean auto-merge, and a content no-op for docs/site/tutorial.md since #55 is a superset). A full docs build was run with the whole pipeline executing for real — remote zarr read, cellSAM segmentation, predict, and both napari screenshots — build succeeded with no cell errors and both images rendering with real content.

That build also confirmed this PR's central premise: napari does render headlessly, so the screenshot cells belong in the source. It needs Xvfb plus a real GLX context (Qt's offscreen platform provides none) and Mesa llvmpipe.

One follow-up on the rebuild note at the end of this PR's description: the from-scratch run surfaced a bug in #55's new guard cell, fixed in 47160ac — int(mask.max()) reported the largest label ID (4644) rather than the cell count (1743), because cellSAM's labels are not contiguous. It only reproduces on a real cellSAM run; the archive's precomputed mask has contiguous labels, so the two numbers agree there.