Native keyboard-focus tests flake with wayland_scope_evidence_expired under coverage #632

Summary

tests/test_computer_native_keyboard_focus_class.py fails intermittently in CI's instrumented coverage-no-drop run: the native observation is rejected with wayland_scope_evidence_expired before the assertion. Each time, the plain tests job passes and a rerun passes, so it is a timing-sensitive test rather than a product failure. It has recurred three times with different parameters.

Occurrences

• v4.9.1 master (2026-09-28): test_predicate_requires_current_native_evidence[backend_label_only] (coverage run; rerun passed).
• campaign/v4.10.0-review @ a5322b47 (2026-09-30, run 36668964361 attempt 1): test_click_then_shortcut_rejects_real_binding_changes[source_digest] (rerun passed).
• v4.11.0 master @ a93348f0 (2026-09-30, run 36671753276 attempt 1): test_predicate_requires_current_native_evidence[geometry]. The assertion expected Image loaded and got {"status": "rejected", "reason": "wayland_scope_evidence_expired", ...} (attempt 2 passed).

Likely cause

The scope-evidence freshness deadline these tests exercise is wall-clock based. Under coverage instrumentation on the shared self-hosted runners, setup can outlast the deadline before the observation, so the evidence expires first.

Expected

The tests control the evidence clock, or measure the deadline from a fixture-owned time source, so instrumentation speed cannot expire it. Keep the production freshness rule unchanged.

Impact

CI only: a master or PR run can fail spuriously and needs a manual rerun, which delays releases.