test(e2e): 40 e2e failures in table/dnd/popover suites — uncovered by CI, pre-existing (not v1.20.1)

[ErickXavier]

Found during post-release verification of v1.20.1 (2026-08-04), running the full Elements e2e suite (cd e2e && npx playwright test) with sibling Core present.

Result: 40 failed / 395 passed, uniform across chromium+firefox+webkit (systematic, not flake):

• table.spec.ts — all 7 tests ×3 browsers (render from state, sorting asc/desc, third-click reset, numeric sort, default sort, sortable class/data-sort-dir, sort indicator)
• dnd.spec.ts — 5 tests ×3 (on:remove / on:receive lifecycle, keyboard Space/Escape drag mode, draggable ARIA)
• popover.spec.ts — ~5 (trigger ARIA before/after open, light dismiss, others)

Bisect: identical 40 failures against Core v1.20.0 and v1.20.1 → NOT a v1.20.1 regression; pre-existing for an unknown number of releases.

Root gap: .github/workflows/ci.yml runs npm test (unit) only — the e2e suite is not exercised by any gate, so these rotted silently. Two work items:

1. Diagnose the 40 failures (genuinely broken plugin behavior vs tests stale against long-changed Core behavior).
2. Add the e2e suite to CI (the Core-clone step already exists in the workflow; needs browsers + the e2e run).

[ErickXavier]

Root-cause analysis complete (code-review pass, 2026-08-04). The reproducible set is 14 tests per browser: 13 stale tests/fixtures, 1 real product bug.

Stale (tests' behaviors still worth covering — update, don't delete):

• table ×7: docs/examples/table/index.html (lines 41-43, 69-71) still uses the sibling <tr else> pattern removed in Core v1.15 — the orphan else row renders permanently (6 rows vs 5) and Core logs the removal warning. Sorting itself works. Migrate fixture to else="templateId" + <template>.
• dnd lifecycle ×2: docs/examples/dnd/index.html still listens on:receive/on:remove/on:reorder; commit 1b109ce (BREAKING, 2026-06-15) renamed events to nojs:dnd-* and updated unit tests but missed this fixture. Payload shapes compatible; rename listeners to on:nojs:dnd-receive etc.
• dnd aria ×3: tests assert aria-grabbed, deliberately removed as deprecated ARIA 1.1 (element.js:246 → aria-roledescription + drag class + live region, commit f072458). Rewrite assertions to the new a11y contract.
• popover aria ×1: test expects aria-haspopup="true"; code correctly sets the more specific "dialog". Update assertion.

Real bug (test legitimately caught it):

• popover light dismiss: .nojs-popover { margin: 0 } + UA inset: 0 paints the popover at viewport (0,0) until the rAF-deferred _positionPopover runs — ≥1 frame mispositioned, swallowing outside clicks (passes solo, fails under load). Fix: position synchronously or keep visibility: hidden until positioned. (src/popover/styles.js:10 + popover.js clickHandler.)

Also: stale test-results/.last-run.json (~380 phantom failures from June 22) should be regenerated; and the CI gap (ci.yml runs Jest only) is what let the June BREAKING rename ship with a broken e2e suite for ~7 weeks.

[ErickXavier]

Resolved in PR #81 (merge b193055), shipped in v1.20.2 (PR #82).

• Real bug fixed: popover painted at viewport (0,0) for ≥1 frame before rAF positioning, swallowing outside clicks — now hidden via inline style until synchronously positioned, exception-safe (try/finally), across all three open paths. Chosen after the CSS-class approach was proven to break dropdowns (menus carry popover="auto" and receive .nojs-popover without the popover open path).
• 13 stale tests/fixtures modernized: table example migrated off the v1.15-removed sibling else pattern; dnd example renamed to nojs:dnd-* events; dnd a11y assertions rewritten to the shipped contract (aria-roledescription + drag class + live-region); popover aria-haspopup="dialog".
• CI gap closed: e2e job added (chromium, node 22, clones+builds Core) — the Playwright suite now gates every PR; first green run: 1m34s. En route we also fixed Core's npm-10-incompatible lockfile (nojs PR #330).

Verification: e2e 435 passed / 0 failed ×3 browsers locally, 870/0 at repeat×2; unit 498/0; CI e2e green on the PR itself. Stale June test-results artifacts were already untracked/gitignored.