Figure capture report — response to FIGURE_CAPTURE_TODO.md

Capture pass run against the local Galaxy (26.2.dev0, worktree history_pages) via Playwright MCP on the dev client (localhost:5173), light theme, 1600×1000 window. All new PNGs are in figures/; source IDs are in figures/MANIFEST.md.

P1 — DONE (all four referenced manuscript figures completed)

How the degenerate “before” was produced (worth knowing for the caption / methods). There was no stored degenerate workflow. The off-graph ATAC notebook is page 36ddb788a0f14eb3; its only content directives are two history_dataset_as_image(...) references, both pointing at re-uploaded PNG datasets (64ed206b5444cf1a “DESeq2 PCA genome-wide”, e1a34e8a9807a4ae “Volcano genome-wide”) — Data-Fetch uploads with no creating tool job. Extracting a workflow seeded by exactly those two HDAs (POST /api/workflows/extract, hda_ids=[…]) yields workflow 4b187121143038ff: two input nodes, zero tools — the literal degenerate result. The 5-step “after” is the on-graph extract 0a248a1f62a0cc04. So the before/after pair is a real, reproducible artifact of the same analysis, not a mock-up. (This also independently confirms the paper’s claim: the off-graph notebook seeds 2 dead inputs / 0 tools; the numbers in the draft match.)

P2 — partial

Revision panel caveat (important for how you caption it). The MRSA page b887d74393f85b6d has 5 revisions; the panel badges them (AI) (= edit_source="agent", 4 of them) and (Unknown) (the initial pre-edit_source revision). It is clean, real, and on-message for agent authorship (evidence layer 4) — an entire scientific notebook authored across four agent edits. BUT it does not show the user + agent + restore mix the TODO asked for ideally, and it is not the UC2 GATA1-null interpretive correction. I checked the candidates:

• No real UC page carries all three edit_source kinds. The UC pages are either all-agent (MRSA: 4×agent) or single-revision (the “clean, extractable” UC2/UC3 finals).
• The full user/agent/restore variety exists only in synthetic test pages (e.g. c385e49b9fe1853c alternates null/agent; the “Restore Test” page 417e33144b294c21 is where restore-source revisions live) — those have no scientific content, so they’d make a weaker figure.
• The GATA1-null correction is described in usecases/UC2_DEBRIEF*.md but is not isolated as a single labeled revision on a clean UC2 page in this server state.

If you want the ideal figure (a real scientific revision showing user→agent→restore), it would need to be staged deliberately: take a clean UC2 page, make a user edit that states the GATA1 claim, an agent edit that corrects it to the GATA1-null framing, then a restore — capturing the panel after. I did not stage this (it would manufacture provenance for a figure). Flagging the decision for you.

Report continuity (not done) — exact procedure if you want it. The claim is “the notebook narrative travels into the extracted workflow’s report.” The extracted workflows here were produced with the page carried in as the report (from_page_id), so the report exists on the workflow. Viewing/screenshotting it cleanly needs either the workflow-report editor or an invocation report render — neither is a one-click read-only view in this build, so I left it. Lowest-friction path: run an invocation of 33b43b4e7093c91f, then GET /api/invocations/{id}/report (or .pdf) and screenshot the rendered report beside uc1_notebook_full.png.

Quality re-captures — NOT done (procedure documented)

The TODO asks to redo the uc*_notebook_full.png shots so the actual UC history (not an empty “Unnamed history”) shows in the right panel. I did not complete these because the page editor’s right-hand panel shows the current user history, not the page’s source history — so it requires switching the session’s current history first, which the dev client does via UI, not a clean REST call. Procedure to finish:

1. In the client, switch current history to the UC history (UC1 48916fac0de9a85d, UC2 96d9e11f37f34b29, UC3 241d84796a24640a) via the history switcher.
2. Open the notebook: /pages/editor?id=<page> (UC1 eafb646da3b7aac5, UC2 42a2c611109e5ed3, UC3 a7e42332dab8f5db).
3. Capture a full-app shot (proves it lives in Galaxy, history populated) and a content-cropped shot (clean figure body). The content body element is the MarkdownEditor preview; crop to it for the clean version.

The existing uc*_notebook_full.png remain usable in the meantime (the only defect is the empty history panel on the right edge).

Notes / gotchas for the next capturer

• Navigating away from the Workflow Editor fires a beforeunload dialog (unsaved-changes guard) — the browser_navigate call times out; accept the dialog with browser_handle_dialog.
• The activity bar floats over the left edge of .workflow-canvas; when cropping, clamp the left edge to ≥ ~44px (image-coords) or zoom out once so the leftmost node clears it.
• Large graphs (29/34 steps) overflow the canvas at default fit — click Zoom Out once or twice, then re-measure the node bounding box before cropping.
• Heatmap datasets are 2000×2000 PNGs; grab them at full res via GET /api/datasets/<id>/display?to_ext=png and downscale, rather than screenshotting the rendered <img>. (The matplotlib title “…by mobility context” is clipped at the right edge in the source PNG — present in both heatmaps, not a capture artifact.)