Restore full hosted E2E matrix after stabilizing macOS baselines

[ashwin-pc]

PR #73 temporarily scopes hosted browser checks to the stable realtime functional specs because the full mobile/tablet/desktop/auth matrix has deterministic and hanging failures on macos-latest that do not reproduce in the local sharded runner.

Track the structural CI fix here:

• classify functional vs visual tests explicitly (tags/projects)
• isolate visual baselines from functional checks
• remove hard-coded local filesystem assumptions from hosted tests
• ensure each Playwright project has bounded startup and teardown
• restore the complete functional matrix as a required PR check

Context: #70, PR #73.

[ashwin-pc]

Additional findings from PR #73 review:

• tests/e2e/retry-errors.spec.ts is independently flaky on both the PR and origin/main (observed 1/8 with --repeat-each=4 --retries=0, racing retrying assistant request vs response failed).
• The first-load service-worker controllerchange reload can destroy a Playwright execution context immediately after goto; consider a shared helper that waits for SW control before interaction.