fix: Wait for app port to bind before navigating (#448)

[schloerke]

Fixes #448

Summary

An intermittent AppDriver$new() failure — Shiny app did not become stable — traces back to a startup race: Shiny prints its Listening on http://… line one statement before it binds the listening socket (rstudio/shiny#4400). app_start_shiny() parsed the URL from that stderr line and let initialization navigate the browser immediately, so on a slow or loaded host the browser could reach a not-yet-bound port, land on Chrome's error page, and when the app finally bound, the error-page recovery navigation raced the readiness/idle checks into the opaque abort (which fires well before load_timeout).

This adds an app_wait_for_serving() helper that polls the port with pingr::is_up() — the same reachability check app_httr2_get() already uses — until it accepts a connection or load_timeout elapses, then returns. Initialization therefore navigates against a bound socket. If the port never comes up in time, the helper returns quietly and startup proceeds as before (graceful fallback). This mirrors the approach proposed by @nbenn in the issue.

Verification

Using the issue's deterministic reproducer (tracing startServer to sleep 5s before bind): without the fix it aborts in ~6s with did not become stable … Caused by error in app_wait_for_idle(); with the fix the app becomes stable in ~8s. New unit tests in tests/testthat/test-app-start.R cover both the unbound-port (returns FALSE, honors timeout) and bound-port (returns TRUE) cases.

[schloerke]

CI status — failures are unrelated to this change; merging via admin

The Package checks workflow is red, but neither failure is caused by this PR's code:

1. test-app-image test-app snapshot failures — pre-existing and reproducible on main. Every scheduled Package checks run on main has failed on these screenshot-comparison snapshots for months (e.g. run 27944676614 on 3844574). They stem from environment/rendering differences, not this change. This PR touches only an internal startup helper, one Suggests entry, and a new unit test — nothing in the image-rendering path.

2. man/AppDriver.Rd \if/\subsection/\code macro warnings (which fail the build under error_on = "warning") — introduced by the CI routine job's devtools::document() auto-commit, which regenerated the docs with roxygen2 8.0.0, whereas the package is documented with 7.3.3 (RoxygenNote: 7.3.3). This is a CI-tooling regression, not a source change here. main's Rd files check OK with 7.3.3.

To keep main's documentation intact, I've dropped the bot's devtools::document() (roxygen 8.0.0) and use_tidy_description() auto-commits from the branch. The PR now contains only the intended change:

• R/app-driver-start.R — wait for the app port to accept a connection before navigating (new app_wait_for_serving() helper)
• tests/testthat/test-app-start.R — unit tests for the helper
• DESCRIPTION — declare httpuv in Suggests (used by the new tests)
• NEWS.md — changelog entry (credits @nbenn)

test-actions (the app-driver integration workflow) passes. Merging with admin since the remaining Package checks failures are pre-existing/environmental and unrelated to this change.