Drive termherd over MCP

The server's instructions and each tool's description carry the calling rules; docs/agents/mcp.md holds the design and the open defects. The loop:

1. Find the target. list_sessions, then address it by handle, never by the Claude resume_id.
2. Act. run_in_session (a raw write, not a bracketed paste) returns at once. For a prompt to Claude, prefer prompt_in_session, which waits for the prompt to start before waiting for it to finish.
3. Wait. wait_for_status for busy first, then for idle: just after typing, the status is still the one from before. Never poll snapshot.
4. Observe. read_terminal for text; screenshot only for render, colour or glyph questions — it weighs hundreds of times a snapshot.

Diagnosing a live defect:

• press_keys tests a binding through the live keymap, run_action the behaviour by action name. An open prompt eats both; escape leaves any prompt. Read each press's outcome (ran, inert + reason, overlay, typed, unbound).
• PTY bytes are the ground truth: in a fresh shell run stty raw -echo; for i in {1..N}; do dd bs=64 count=1 | xxd -p; done, then press.
• Clipboard: printf marker | pbcopy (outside the sandbox), act, pbpaste.
• mouse_in_session answering forwarded for a move proves the child asked for DECSET 1003.
• A synthesised key has key == text: Option-composed characters, and keys press_keys cannot name, need the user's hands.