test(tui): probes that show what flags and settings do to the Python agent - #57

zypi ──ACP──> wire_log_agent.py ──ACP──> context_spy_agent.py (Python agent)
                  │ wire.jsonl                │ spy.jsonl: model, reasoning, tools, system prompt
                  └──────────── wire_probe.summarize / changes(base, run) ────────────┘

scripts/tui_pty/
├── flags_probe.py          # 63 cases: 62 flags (hidden ones that touch the wire or a file too) + a repeat of the base
├── settings_probe.py       # the 37 rows, the permission question, the cursor row, [models].default
├── wire_probe.py           # shared: execute() runs one zypi; summarize() / changes() compare runs
├── wire_log_agent.py       # stand-in agent: copies ACP both ways into a log (stdlib only)
├── context_spy_agent.py    # the Python agent with each LLM call recorded + a scripted tool call
└── README.md               # two table rows, two run lines
scripts/tests/test_tui_probes.py   # 16 tests: the helpers and the stand-in agent (no zypi, no pty)

remember_tool_approvals   unset | true | false : 2 questions each, answers: [Allow once, No, reject]
                          -> no "Always allow ..." row exists; the setting has nothing to switch
default_selected_permission   unset / always_allow_all_sessions / allow_once / allow_command_always
                          -> cursor on "Allow once"   (the page says "Always allow on all sessions")
                          reject -> cursor on "No, reject"   (the only value that does something)
[models] default = mock-b label after the turn: mock-a; LLM call used: mock/mock-a
                          -> the "Default model" row switches this session; the next start ignores it
hunk_tracker_mode, fork_secondary_model, prompt_suggestions, remember_tool_approvals,
[cli] auto_update, [toolset.ask_user_question] timeout_enabled, all set at once
                          -> NO CHANGE on the wire, in the agent's variables, in the LLM call
tools the LLM is offered  bash edit fetch_url find goal_complete goal_update grep ls read web_search workflow write
                          -> no ask_user_question tool