fix(auto-approve): escalate to manual keyboard when auto-Yes doesn't clear the prompt

[Time4Mind]

Problem

auto_approve=on sent the N. Yes digit + Enter into the pane on every poll but never verified the prompt actually cleared. A sub-agent (Agent/Task) WebFetch/WebSearch approval doesn't accept the top-level number hotkey, so the keystroke landed on a busy REPL, the prompt survived, and the poller re-fired the same auto-Yes forever — while _maybe_auto_approve returning True each time shadowed the kb-mode keyboard / bg-status ❓. Net effect: the session wedged with no manual escape hatch (only a raw tmux Esc cleared it).

Diagnosed on a live session (region weather): a web-researcher sub-agent hung ~8 min on unreachable external web; Auto-approved opt=1 spammed the log; stall_finalize fired; the user had to interrupt by hand.

Fix (cross-poll verification + escalation)

• Each poll compares the prompt's signature (extract_interactive_content body+name, fallback to the numbered option block) to the last one auto-approved. Same signature next poll ⇒ the keystroke did not clear it.
• After AUTO_APPROVE_MAX_ATTEMPTS (3) consecutive no-progress attempts on an identical prompt, stop auto-approving and return False → the caller surfaces the manual keyboard (active session) or the ❓ badge (background).
• Distinct prompts (different domain/file) have distinct signatures ⇒ the counter resets, so normal one-shot approvals are untouched.
• _reconcile_no_ui_state pops the counter once the prompt clears (new window_id param).

Log on give-up: auto_approve giving up after N attempts (prompt not clearing) … — surfacing manual keyboard.

Tests

• test_auto_approve_escalates_after_max_attempts
• test_auto_approve_distinct_prompts_never_escalate
• test_auto_approve_counter_resets_when_prompt_clears
• Existing kb-mode debounce tests updated for the new _reconcile_no_ui_state(window_id) signature.

pytest -k "status or interactive or kb_mode" → 67 passed. ruff clean, pyright 0 errors.

🤖 Generated with Claude Code