Add an R Playwright e2e suite

Split out of #185, which closed the R unit-test gaps. This is the one item left: R has no end-to-end tests.

Why it is its own issue

The blocker is fixture infrastructure, not test cases. pkg-py/tests/playwright/ stands up a Python Shiny app per fixture; an R suite needs an equivalent runner for R apps (shiny::runApp() in a subprocess, port allocation, readiness polling, teardown) before a single assertion can be written. That is a different kind of change from writing tests, so it did not belong in #185's PR.

What to cover

The JS bundle is byte-identical between the two packages (md5 335334fd88165d6906de2fd20c2ee174), so an R suite is not re-testing the client — it validates the R server bindings. A subset of the Python suite is enough. Python currently covers:

• module namespaces
• module dependency injection
• bookmark restore
• input-handler types
• ShinyOutput

Bookmark restore and input-handler types are the highest value in R, since those are where the two servers' implementations actually differ (pkg-r/R/bookmark.R uses pinned shiny::: internals; default_input_handler() exists specifically to undo jsonlite/shiny simplification that Python never does).

Module dependency injection is likely not applicable — set_react_page()'s renderer-dependency harvesting is Shiny Express-only and has no R analogue by an explicit decision.

Notes

• decisions/2026-03-17-playwright-testing-architecture.md records the recommended approach (code-gen from TypeScript) for browser testing generally; worth reading before picking a direction.
• The Makefile pattern to mirror: py-test-e2e clears the addopts that otherwise ignore the playwright/ subtree, keeping py-check-tests fast.

Read decisions/2026-03-17-playwright-testing-architecture.md and compare pkg-py/tests/playwright/ with the Makefile's py-test-e2e pattern. Then inspect pkg-r/R/bookmark.R and the existing Python cases for bookmark restore and input-handler types. Done means an R app runner supports subprocess startup, port allocation, readiness polling, and teardown, with a focused R Playwright suite covering the highest-value server bindings.