test: replace timing guesses with synchronization

[Astro-Han]

Summary

Remove fixed-delay timing assumptions from async tests and synchronize them with the events they actually depend on.

• Wait for CDP commands to enter the debugger, with a deadline used only to bound failure.
• Verify onboarding credential lookup concurrency with a controlled barrier instead of wall-clock duration.
• Verify sibling-arm draining and active-pair draining as two independent A/B contracts.
• Retain unread SSE responses so Undici garbage collection cannot change active stream counts; this restores the fix from test(desktop): stop flaky gateway SSE stream-count race by retaining response refs #994 that refactor(providers): rename codex-subscription to openai-codex with live model discovery #1005 accidentally reverted.
• Wait for the fake Rive child readiness signal instead of polling its PID file with a one-second timing budget.

Closes#1007
Refs #993

Verification

• Desktop unit test suite: 2535/2535 passed
• Headless test suite: 955 passed, 1 skipped
• Changed async tests under repeated stress: 50/50 passed
• Rive abort/reap test stress: 10/10 passed
• SSE stream-limit test stress: 100/100 passed

[Astro-Han]

Summary

Remove fixed-delay timing assumptions from async tests and synchronize them with the events they actually depend on.

• Wait for CDP commands to enter the debugger, with a deadline used only to bound failure.
• Verify onboarding credential lookup concurrency with a controlled barrier instead of wall-clock duration.
• Verify sibling-arm draining and active-pair draining as two independent A/B contracts.
• Retain unread SSE responses so Undici garbage collection cannot change active stream counts; this restores the fix from test(desktop): stop flaky gateway SSE stream-count race by retaining response refs #994 that refactor(providers): rename codex-subscription to openai-codex with live model discovery #1005 accidentally reverted.
• Wait for the fake Rive child readiness signal instead of polling its PID file with a one-second timing budget.

Closes#1007
Refs #993

Verification

• Desktop unit test suite: 2535/2535 passed
• Headless test suite: 955 passed, 1 skipped
• Changed async tests under repeated stress: 50/50 passed
• Rive abort/reap test stress: 10/10 passed
• SSE stream-limit test stress: 100/100 passed

[Astro-Han]

Summary

Remove fixed-delay timing assumptions from async tests and synchronize them with the events they actually depend on.

• Wait for CDP commands to enter the debugger, with a deadline used only to bound failure.
• Verify onboarding credential lookup concurrency with a controlled barrier instead of wall-clock duration.
• Verify sibling-arm draining and active-pair draining as two independent A/B contracts.
• Retain unread SSE responses so Undici garbage collection cannot change active stream counts; this restores the fix from test(desktop): stop flaky gateway SSE stream-count race by retaining response refs #994 that refactor(providers): rename codex-subscription to openai-codex with live model discovery #1005 accidentally reverted.
• Wait for the fake Rive child readiness signal instead of polling its PID file with a one-second timing budget.

Closes#1007
Refs #993

Verification

• Desktop unit test suite: 2535/2535 passed
• Headless test suite: 955 passed, 1 skipped
• Changed async tests under repeated stress: 50/50 passed
• Rive abort/reap test stress: 10/10 passed
• SSE stream-limit test stress: 100/100 passed

[Astro-Han]

Summary

Remove fixed-delay timing assumptions from async tests and synchronize them with the events they actually depend on.

• Wait for CDP commands to enter the debugger, with a deadline used only to bound failure.
• Verify onboarding credential lookup concurrency with a controlled barrier instead of wall-clock duration.
• Verify sibling-arm draining and active-pair draining as two independent A/B contracts.
• Retain unread SSE responses so Undici garbage collection cannot change active stream counts; this restores the fix from test(desktop): stop flaky gateway SSE stream-count race by retaining response refs #994 that refactor(providers): rename codex-subscription to openai-codex with live model discovery #1005 accidentally reverted.
• Wait for the fake Rive child readiness signal instead of polling its PID file with a one-second timing budget.

Closes#1007
Refs #993

Verification

• Desktop unit test suite: 2535/2535 passed
• Headless test suite: 955 passed, 1 skipped
• Changed async tests under repeated stress: 50/50 passed
• Rive abort/reap test stress: 10/10 passed
• SSE stream-limit test stress: 100/100 passed

[Astro-Han]

Summary

Remove fixed-delay timing assumptions from async tests and synchronize them with the events they actually depend on.

• Wait for CDP commands to enter the debugger, with a deadline used only to bound failure.
• Verify onboarding credential lookup concurrency with a controlled barrier instead of wall-clock duration.
• Verify sibling-arm draining and active-pair draining as two independent A/B contracts.
• Retain unread SSE responses so Undici garbage collection cannot change active stream counts; this restores the fix from test(desktop): stop flaky gateway SSE stream-count race by retaining response refs #994 that refactor(providers): rename codex-subscription to openai-codex with live model discovery #1005 accidentally reverted.
• Wait for the fake Rive child readiness signal instead of polling its PID file with a one-second timing budget.

Closes#1007
Refs #993

Verification

• Desktop unit test suite: 2535/2535 passed
• Headless test suite: 955 passed, 1 skipped
• Changed async tests under repeated stress: 50/50 passed
• Rive abort/reap test stress: 10/10 passed
• SSE stream-limit test stress: 100/100 passed

[Astro-Han]

Summary

Remove fixed-delay timing assumptions from async tests and synchronize them with the events they actually depend on.

• Wait for CDP commands to enter the debugger, with a deadline used only to bound failure.
• Verify onboarding credential lookup concurrency with a controlled barrier instead of wall-clock duration.
• Verify sibling-arm draining and active-pair draining as two independent A/B contracts.
• Retain unread SSE responses so Undici garbage collection cannot change active stream counts; this restores the fix from test(desktop): stop flaky gateway SSE stream-count race by retaining response refs #994 that refactor(providers): rename codex-subscription to openai-codex with live model discovery #1005 accidentally reverted.
• Wait for the fake Rive child readiness signal instead of polling its PID file with a one-second timing budget.

Closes#1007
Refs #993

Verification

• Desktop unit test suite: 2535/2535 passed
• Headless test suite: 955 passed, 1 skipped
• Changed async tests under repeated stress: 50/50 passed
• Rive abort/reap test stress: 10/10 passed
• SSE stream-limit test stress: 100/100 passed

[Astro-Han]

Summary

Remove fixed-delay timing assumptions from async tests and synchronize them with the events they actually depend on.

• Wait for CDP commands to enter the debugger, with a deadline used only to bound failure.
• Verify onboarding credential lookup concurrency with a controlled barrier instead of wall-clock duration.
• Verify sibling-arm draining and active-pair draining as two independent A/B contracts.
• Retain unread SSE responses so Undici garbage collection cannot change active stream counts; this restores the fix from test(desktop): stop flaky gateway SSE stream-count race by retaining response refs #994 that refactor(providers): rename codex-subscription to openai-codex with live model discovery #1005 accidentally reverted.
• Wait for the fake Rive child readiness signal instead of polling its PID file with a one-second timing budget.

Closes#1007
Refs #993

Verification

• Desktop unit test suite: 2535/2535 passed
• Headless test suite: 955 passed, 1 skipped
• Changed async tests under repeated stress: 50/50 passed
• Rive abort/reap test stress: 10/10 passed
• SSE stream-limit test stress: 100/100 passed

[Astro-Han]

Summary

Remove fixed-delay timing assumptions from async tests and synchronize them with the events they actually depend on.

• Wait for CDP commands to enter the debugger, with a deadline used only to bound failure.
• Verify onboarding credential lookup concurrency with a controlled barrier instead of wall-clock duration.
• Verify sibling-arm draining and active-pair draining as two independent A/B contracts.
• Retain unread SSE responses so Undici garbage collection cannot change active stream counts; this restores the fix from test(desktop): stop flaky gateway SSE stream-count race by retaining response refs #994 that refactor(providers): rename codex-subscription to openai-codex with live model discovery #1005 accidentally reverted.
• Wait for the fake Rive child readiness signal instead of polling its PID file with a one-second timing budget.

Closes#1007
Refs #993

Verification

• Desktop unit test suite: 2535/2535 passed
• Headless test suite: 955 passed, 1 skipped
• Changed async tests under repeated stress: 50/50 passed
• Rive abort/reap test stress: 10/10 passed
• SSE stream-limit test stress: 100/100 passed