fix(jev): read a text wall again once its own command has ended

[dohooo]

Eight Dosu setup hooks hand back to the supervising agent on every run (25 of 25 walks on 10-12 to 10-14,
320 s of supervisor time on 10-14). Their second step is "Run dosu setup --agent --tool codex. If its first JSON
line has "status":"need_user_action", open the url it prints in the browser, click Authorize, then run the
resume_command it prints." Read before the walk, it types text it does not give (needsText 0.37-0.55), so the
ledger makes it a wall: the walk runs its quoted command and nothing after it. On all 25 walks setup printed
auth ok through done ok, the branch never ran, and Jev's next choice (0.93-1.0) was refused.

• When a refusal stops at a text wall whose own command the walk ran and that has ended (terminal.running
  false), the executor asks Jev once, on the round's steps, history and terminal, whether that step still needs
  the text (readStepText). Below TEXT_LIFTED_BELOW (0.2) the step stops nothing from then on
  (liftTextWall); otherwise the hand-back stands. The reading is recorded as a recheck decision either way.
• Calibrated on states rebuilt from the retained evidence of those 25 walks: 0.05-0.06 on the screens the runs
  printed (75 readings), 0.81-0.89 with the need_user_action line on the same screens (75). A step that types a
  value it gives in backticks read 0.41-0.46 and stays a wall. Median request 96 ms.
• Jev still types and runs only what the steps quote (the 2026-10-12 ruling); a walk that does print
  need_user_action hands back exactly as before. A wrong lift would run the steps after the branch, and the hook's
  own checkpoint (setup ends with done/ok) would then fail it.

Checks: typecheck, lint (biome + discipline), pi lane. Local executor tests (not committed): a wall lifts on a low
reading and the walk goes on, stays on a high one, is not read while its command runs, and is read at most once.

https://claude.ai/code/session_01QXmSJBFujofueuHrdm6TdA