Two tests turning good pulls red fail on and off on main: make them steady or quarantine them today #8995

Both tests now wait on what they need, and neither is quarantined. Both files passed 10 of 10 runs while 8 CPU-burning processes ran alongside, and tsc --noEmit is clean. I did not run scripts/gate.ts, design-lint or the rest of the suite; the forge runs the full check on this push. Only test files changed.

test/tui-backups.test.ts, "a bad endpoint is refused…"

• Cause: the failure was in the test, not the code. The prompt shows only the start of the refusal ("The endpoint must be an http(s) URL without crede…"). The full sentence sits on the screen behind it, and the prompt covers the middle of that line, so asserting the whole sentence while the prompt was open could never pass.
• Fix: while the prompt is open, the test checks for the visible start of the refusal and 1/5. After esc it checks for the full sentence on the page behind. That is also what the prompt's hint, "a refusal is said whole behind", promises.

test/conversation-queue.test.ts, "Stop forgets what waited…" and "steer stops the turn…"

• Cause: the fake agent writes its log line before the forge has read its partial answer. The tests treated that line as "the turn is running" and stopped or steered in that gap, so the partial answer was sometimes lost. The until helper also gave up silently after 4 seconds and let the test carry on into a wrong answer.
• Fix:

- Both tests now wait until the first partial answer is stored before they stop or steer. - until now throws a named timeout, with a 10-second limit. A hang fails with a message instead of a later wrong assertion.

I could not reproduce the conversation-queue flake in the sandbox before the change, so its proof is the race read from the code plus the 10 clean runs, not a before/after. The steadiness proof is the 10 runs under load, as the flake rule asks. Pulls parked by these two tests can be sent back.

Closes #8993

Both tests now wait on what they need, and neither is quarantined. Both files passed 10 of 10 runs while 8 CPU-burning processes ran alongside, and tsc --noEmit is clean. I did not run scripts/gate.ts, design-lint or the rest of the suite; the forge runs the full check on this push. Only test files changed.

test/tui-backups.test.ts, "a bad endpoint is refused…"

• Cause: the failure was in the test, not the code. The prompt shows only the start of the refusal ("The endpoint must be an http(s) URL without crede…"). The full sentence sits on the screen behind it, and the prompt covers the middle of that line, so asserting the whole sentence while the prompt was open could never pass.
• Fix: while the prompt is open, the test checks for the visible start of the refusal and 1/5. After esc it checks for the full sentence on the page behind. That is also what the prompt's hint, "a refusal is said whole behind", promises.

test/conversation-queue.test.ts, "Stop forgets what waited…" and "steer stops the turn…"

• Cause: the fake agent writes its log line before the forge has read its partial answer. The tests treated that line as "the turn is running" and stopped or steered in that gap, so the partial answer was sometimes lost. The until helper also gave up silently after 4 seconds and let the test carry on into a wrong answer.
• Fix:

- Both tests now wait until the first partial answer is stored before they stop or steer. - until now throws a named timeout, with a 10-second limit. A hang fails with a message instead of a later wrong assertion.

I could not reproduce the conversation-queue flake in the sandbox before the change, so its proof is the race read from the code plus the 10 clean runs, not a before/after. The steadiness proof is the 10 runs under load, as the flake rule asks. Pulls parked by these two tests can be sent back.

Left #8993 open: 1 unticked item remain (Say how each got past the admission rule. Pulls parked by these are sent back.).