[py] run the BiDi tests against the Grid server

[AutomatedTester]

🔗 Related Issues

Part of #17978 (the cross-binding half of that gap remains open).
Uncovered #17979 along the way.

💥 What does this PR do?

Runs the Python BiDi tests against Grid, which no binding did before.

BIDI_TESTS was excluded from every remote suite glob, so there was no test-<browser>-remote-bidi target and the path where the client speaks BiDi over Grid's websocket proxy had never been exercised from Python. No client or conftest.py change was needed — --remote and --bidi are independent code paths that already compose, LocalNode already rewrites webSocketUrl to /session/<id>/se/bidi, and ProxyNodeWebsockets already tunnels it. The missing piece was purely the Bazel target.

Result across all 26 BiDi targets, chrome and firefox: zero remote-specific failures. Every failure observed reproduced identically without --remote, so no xfail_remote markers were added — marking them would have mislabelled a headless/timing problem as a Grid problem. The two genuine ones are fixed instead:

• test_activate_browsing_context asserted a focus handover that happens asynchronously, read exactly once. Now polled. A plain wait is not sufficient on its own: headless Chromium has no window manager and never hands focus over at all (a 5s wait timed out 5/5), while headless Firefox does. So it skips only when the wait expires and the run is headless — headless Firefox keeps the full assertions, and headed runs still fail loudly on a real regression because the skip cannot fire outside headless.
• test_wheel_scroll_with_duration spread a 100px scroll over 500ms then asserted pageYOffset == 100 immediately; Firefox was reliably one pixel short. Now waits for the scroll to settle, keeping the exact assertion rather than widening it to a tolerance. Failed 3/3 runs before, 0/3 after.

Also included, all pre-existing inconsistencies found while building the above:

• Remote suites carried tags = ["no-sandbox", "remote", "<browser>-remote"], omitting the COMMON_TAGS + bare browser name every non-remote suite has. The tag that mattered is the browser name: py/TESTING.md documents --test_tag_filters=chrome, and that matched no remote target. Ruby already does this and says why in a comment.
• Remote suites omitted target_compatible_with, which the bidi and per-browser suites pass through.
• conftest.py repeated the "skip unless --remote" guard three times and the xfail_<driver>/xfail_remote marker dance twice.

🔧 Implementation Notes

The new suite is deliberately not folded into the test-<browser>-remote aggregate, and gets its own remote-bidi-tests CI job, so a red run in new coverage cannot redden the established remote job. Easy to merge in later once it has proven itself.

The two duplicated marker blocks in conftest.py had drifted apart, so unifying them fixes two latent bugs on the clean_driver path: it did not support the condition kwarg and passed it straight into pytest.xfail(), which rejects it (TypeError); and neither block popped run before calling pytest.xfail(), so a documented run=True raised TypeError too. The yield/return after pytest.skip() in both was dead code, since pytest.skip() raises. Verified behaviour-identical against alerts_tests, which exercises all three marker paths (it carries condition=sys.platform == "darwin", run=False plus plain xfails): 20 passed, 1 skipped, 2 xfailed, before and after.

The COMMON_TAGS omission was incidental rather than deliberate — the per-browser remote targets in #16851 inherited the minimal tag set of the single test-remote target they replaced, and no rationale for the choice is recorded. One behaviour change to be aware of: --test_tag_filters=chrome now also matches remote targets, which is slower but is what the docs imply should happen.

skip-rbe was considered and deliberately not added. These targets are selected by the RBE run (.skipped-tests is empty and -skip-rbe is the only filter), but CI - RBE on trunk is green and Ruby has run remote tests on RBE since #13906, so the Grid-on-RBE path works.

Each concern is a separate commit if that is easier to review.

🤖 AI assistance

• No substantial AI assistance used
• AI assisted (complete below)
  • Tool(s): Claude Code
  • What was generated: the Bazel targets, the CI job, the conftest.py helpers, the two test fixes, and the doc updates, plus the verification runs behind the claims above
  • I reviewed all AI output and can explain the change

💡 Additional Considerations

Verified on macOS arm64 with pinned browsers, headed and headless, chrome and firefox. The new remote-bidi-tests job has not yet run on ubuntu/Xvfb, so its first CI run is the real test of the headless reasoning above — CI is headed under fluxbox, where the chrome skip should not fire and the full assertions should run.

Follow-ups, not in this PR:

• _start_bidi passes only (url, timeout, interval) to WebSocketConnection, dropping ClientConfig's username/password/auth_type, extra_headers and ca_certs (wss:// hardcodes cert_reqs: CERT_NONE) plus websocket_max_message_size. Not reachable through the local standalone used here (plain http, no auth), so it is out of scope, but it would break an authenticated or TLS Grid. It wants to stack on [py]: Apply ClientConfig user_agent and extra_headers per connection #17944, which reworks how per-connection headers are resolved.
• The remaining bindings in [🚀 Feature]: No binding runs its BiDi tests against Grid #17978.
• The firefox flakiness in [🐛 Bug]: Flaky Firefox failures in py _bidi input tests #17979.

🔄 Types of changes

• Cleanup (formatting, renaming)
• Bug fix (backwards compatible)
• New feature (non-breaking change which adds functionality and tests!)

🤖 Generated with Claude Code

Co-Authored-By: Copse noreply@copse.dev
Copse-Models: acp:claude-agent-acp#opus[1m]