Test gaps after the browser suite landed: picker paths, node-grid surfaces, PIL-stub blind spot

[laurigates]

Re-scoped 2026-08-16. §1 is closed — the browser suite it asked for landed in #107. What follows is what actually survives, re-verified against HEAD rather than carried over.

Measured at f819981: pytest 296, vitest 147 (9 files), Playwright 17.

§1 — CLOSED by #107 (f2ed146)

tests/e2e/ exists: stdlib-only stub server (server.mjs) serving the built web/dist/index.js at its real extension URL plus /gallery_loader/{base,list,thumb,file,pins} with real PNG bytes, a fixture page (fixture.html) driving the bundle's exported openImagePicker(), and probe.js — the scrollTop-setter spy that separates "clamped at assignment" from "moved afterwards". Chromium at a 390×844 phone viewport, its own CI job.

Both features that were scoped around the absence are back:

• Scroll restore / per-directory memory is implemented, through the kit's shared installScrollRestore, and pinned by tests/mutations-e2e.json (6 real mutations + a CONTROL).
• First-open centring works in flat view, which was the thing a single bare scrollTop write could not do.

Still true, and worth keeping on the record: Chromium-only, so iOS momentum scrolling is uncovered by any tier.

The one item from §1 that was not closed is the measurement gap, so it moves down into §2:

tests/js/image-picker.test.js asserts which object the IntersectionObserver is rooted on. comfyui-image-browser additionally measures the consequence — 400/400 off-screen cards intersect with the grid as root vs 20/400 with the real scroller. Ours would still pass if the root were right and the behaviour wrong.

tests/e2e/ can now express that (REGRESSION N already asserts only the restored band's thumbnails are fetched), so it is a browser-tier item now, not a blocked one.

§2 — Picker paths still untested

Re-checked against src/image-picker.ts at HEAD; all still uncovered:

• Directory mode — the footer "Use this folder" commit (commitFolder, incl. the . sentinel at a root), and that file cards carry is-inert there.
• VHS path mode — the absolute-path commit via joinAbs, and that state.absPath seeds from /base when the widget is empty.
• Navigation — navigateInto / navigateUp, breadcrumb clicks (data-sub vs data-abs), and a tab switch resetting subfolder.
• .__none__ sentinel from the frontend side. The backend has a test named for the trap (test_directory_mode_sentinel_still_lists_nothing); nothing asserts the picker actually sends it in directory mode.
• Lazy-load consequence, not just the observer root (moved from §1) — the 400/400-vs-20/400 measurement, now a tests/e2e/ item.

§3 — Node grid

• The commit-contract divergence — done in test(node-grid): pin the commit contract on the surface that had none #111. Also resolved as a question: it is not one contract implemented twice, it is two widgets. The picker writes core LoadImage's native combo, whose options are sorted(files) off the input dir with no annotation (ComfyUI/nodes.py, LoadImage.INPUT_TYPES), so bare matches an existing option and pre-existing workflows do not churn. The node grid writes GalleryLoadImage's own STRING widget, which has no option list, so annotating states the root outright instead of leaning on parseAnnotated's bare-relative fallback. Both resolve identically through get_annotated_filepath. Now documented in CLAUDE.md § Value contract and pinned on both sides with 4 mutations + a CONTROL.
• Type-chip switching beyond what test(node-grid): pin the commit contract on the surface that had none #111 needed (it drives the output chip, so this is now partial).
• The .gl-pathinput free-text absolute-path field (Enter and blur commit paths).
• DOM-widget mounting and the hideOnZoom teardown.

§4 — Backend

• Unchanged and still correct: conftest.py stubs PIL with MagicMocks, so _scan_file_entry's width/height probe always raises into its own except and dimensions come back None. Nothing asserts dimensions, which is right today but means a regression that broke the probe outright would be invisible. A narrow real-PIL test (or a stub returning a size) closes it.

Adjacent but tracked in #14, not here: the /thumb + /file handler-level coverage, which #112 adds.

---

Original body preserved below for provenance.

Original issue text (2026-08, pre-#107)

Follow-up to #85. That PR added the pack's first real DOM coverage (38 → 68 vitest specs, plus the first-ever tests for gallery_loader.ts), but it also surfaced gaps it could not close — and in two places I scoped work down because of them. Filing so the reasons don't evaporate.

1. There is no browser suite, and it is load-bearing

tests/js/ is jsdom, which performs no layout. It accepts scrollTop = 500 on a zero-height scroller and reads back 500 — detached or not. So an entire class of bug is structurally invisible here, and the sibling pack has a Playwright suite (comfyui-image-browser/tests/e2e/) specifically because that class was live there twice.

Two decisions in #85 were made around this absence rather than on the merits:

• scrollToSelected is guarded off in flat view. Its single bare scrollTop write lands against a grid whose thumbnails are all still data-src placeholders, so a real engine clamps it at the instant of assignment and the view settles somewhere arbitrary. The fix is a bounded re-assert loop — but a re-assert loop that no test can falsify is worse than not scrolling, so flat view simply doesn't scroll-to-selected.
• Scroll restore / per-directory memory was not ported at all, for the same reason.

There is also a measurement gap the structural tests can't cover (see §2 above, where it now lives).

2. Picker paths still untested

(unchanged — carried up)

3. Node grid coverage is a beachhead, not a suite

(carried up; the commit-contract bullet is now done)

4. Backend

(carried up unchanged)

[laurigates]

Re-scoped against HEAD (f819981). Two changes worth calling out rather than burying in the edit:

• §1 is closed by feat(picker): remember and restore the scroll position per folder #107 — tests/e2e/ exists, and both features that were scoped around the missing browser tier (scroll restore, flat-view centring) are implemented and pinned by tests/mutations-e2e.json. The one part of §1 that was not closed — measuring the lazy-load consequence rather than the observer root — moved down into §2, because the browser tier can now express it.
• §3's first bullet is done in test(node-grid): pin the commit contract on the surface that had none #111, and it turned out not to be a bug. The divergence is two widgets, not one contract implemented twice: the picker writes core LoadImage's combo (options are sorted(files), bare, so a bare value matches one and workflows don't churn), the node grid writes GalleryLoadImage's STRING widget (no option list, so the annotation is free). Both resolve through get_annotated_filepath. Now documented in CLAUDE.md and pinned on both sides.

Current counts, measured: pytest 296, vitest 147 (9 files), Playwright 17.

[laurigates]

Re-verified against origin/main (bc4aaab).

§1 confirmed closed

tests/e2e/ exists and holds six files, not the five named in the re-scope — harness.js was omitted there:

tests/e2e/fixture.html
tests/e2e/harness.js
tests/e2e/playwright.config.js
tests/e2e/probe.js
tests/e2e/scroll-restore.spec.js
tests/e2e/server.mjs

npx playwright test --list -c tests/e2e/playwright.config.js → 17 tests in 1 file, project chromium-mobile. The last one listed is REGRESSION N — only the restored band's thumbnails are fetched, which is the assertion §2's moved-down lazy-load item was waiting on a browser tier to be able to express.

Still true: Chromium-only, so iOS momentum scrolling remains uncovered by any tier.

Counts corrected

The body records Measured at f819981: pytest 296, vitest 147 (9 files), Playwright 17. Measured at bc4aaab, by running the suites rather than counting statically:

Tier	Recorded	Measured now	Command
pytest	296	318 collected	uv run pytest --collect-only -q
vitest	147 (9 files)	152 in 9 files	npx vitest run
Playwright	17	17 in 1 file	npx playwright test --list

The pytest figure is collected tests, which is above the 259 def test_ functions across the 11 tests/test_*.py files because four of them use parametrize. Worth stating which of the two a future count means — the previous one did not, and the two numbers differ by 59.

Measured in a detached worktree at origin/main, not in a working tree that was sitting on a feature branch with 3 test files ahead.

Unchanged

§2, §3 (minus the #111 item) and §4 re-read at HEAD and still stand as written. No re-scope beyond marking §1 done and correcting the two drifted counts.

Metadata: assignee (laurigates) and type label (chore) were already present — nothing backfilled.