ci(hub-client-e2e): fix Playwright install hang (Node pin) + smoke-all seed/sync flakiness

[gordonwoodhull]

Fixes two independent problems that were keeping the Hub-Client E2E workflow red. CI is now green (run 26772812401: 78 passed / 14 flaky / 0 failed functional, 6 visual passed, ~13 min with 2 workers).

1. Install hang (Node + Playwright)

The Install Playwright step hung until the 6 h job cap (every run since ~2026-05-27). Root cause: a yauzl extraction-hang regression in Node ≥ 24.16 — after the browser zip downloads to 100 %, extraction deadlocks. It affects Playwright < 1.60.0 (we pin @playwright/test at ^1.50.0 → 1.58.0), and our CI runs node-version: '24' (floats past 24.16). Host-independent; not the CDN. Refs microsoft/playwright#40724, fixed by #40747 in PW 1.60.0.

Fix: pin Node to 24.15.0 (last release before the regression). Keeps Playwright 1.58 / Chromium 145, so visual baselines are unchanged. This is a stopgap — the durable fix is upgrading to PW ≥ 1.60 and dropping the pin (tracked in bd-2njja; deferred because 1.60 bumps Chromium 145 → 148 and needs a visual-baseline regen).

2. smoke-all flakiness (seed-before-synced race)

Unblocking the install surfaced pre-existing smoke-all flakiness (same signature on the 2026-05-27 run; not caused by the Node pin). Root cause is a test-harness race, not the render pipeline, not auth (the 401 on /auth/me is the benign anonymous response), and not parallelism (verified: workers: 1 did not help):

seedProjectInBrowser → addProject writes IDB only. A seeded project reaches the synced project set only via reconcileIntoConnectedProjectSet, which the app runs on the status→connected transition and which requires isConnected(). The test seeds after the set is already connected, so that effect never re-fires — the seed lands in the set only if a fortuitous WS reconnect re-triggers reconcile. When it doesn't, the project is absent from the set, navigation to /p/<id> drops to the landing page, and waitForPreviewRender times out at 75 s. Confirmed via the failing run's trace + final screenshot ("No projects yet", empty set).

Fix (keeps the full VFS → Automerge → WASM → preview path end-to-end — just stops racing it): expose the live projectSetService + the idempotent reconciler on window.__quartoTest, and have seedProjectInBrowser wait for a real peer connection, run the reconciler, and wait until the project is observably present in the connected set before returning. Bounded 30 s waits fail loudly if the sync server is truly unreachable. workers stays at 2.

Rejected the alternative of seeding content locally — it would bypass the very integration smoke-all exists to exercise.

Follow-ups (tracked)

• bd-2njja — upgrade @playwright/test ≥ 1.60.0, drop the Node pin, regen Chromium-148 visual baselines.
• bd-3nzyd — residual smoke-all flakiness: 14 tests still pass-on-retry (likely project file-doc sync during render, a different doc than the project-set race fixed here); harden similarly.

Note on main

Two earlier direct-to-main commits (5f16330f, 78f563b7) tried CDN-based install fixes that didn't work; this PR restores the pristine workflow and supersedes them.