Clear TODO backlog: --debug/--profile, codex selection, MCP orientation, perf

- Add --debug[=DIR] / --profile[=DIR] flags that write run artefacts
  (patterm.log, events.jsonl, per-child raw PTY captures, CPU + heap
  + goroutine pprof) to a dir without polluting stdout/stderr.
- Strengthen vendor-TUI orientation in three places (MCP
  initialize.instructions, the spawn_agent tool description, and
  help('spawning')) to head off codex's habits of poking the Unix
  socket via perl and shelling out to launch peers — both bypass
  caller identity and produce orphaned top-level tabs.
- Fix click-and-drag text selection from alt-screen TUIs. Host SGR
  mouse reporting now follows the focused child's screen side
  instead of being permanently armed; alt-screen TUIs that need
  mouse re-enable it themselves and the toggle is forwarded.
- Move drawSidebar() off the per-PTY-chunk hot path. Long claude
  session resume was paying a full sidebar rebuild for every
  scrolled chunk; the chrome ticker now drains a dirty flag at 60 Hz.
- Gate the per-chunk Title() CGO poll on a containsOSC scan so
  codex/ratatui's many SGR-only chunks no longer pay a CGO call each.

internal/mcp/protocol.go

@@ -27,6 +27,24 @@ var serverInfo = map[string]any{
 	"version": "0.1.0",
 }
 
+// serverInstructions is returned in the MCP `initialize` response. MCP
+// clients show this to the underlying LLM as context for how to use
+// the server. Failure modes we've seen and want to head off:
+//   - The agent assumes patterm is something it has to launch (running
+//     `patterm` or `patterm mcp-stdio` from its own shell). It's
+//     already attached — it just calls the tools.
+//   - The agent reaches for shell tools (perl / nc / socat / curl) to
+//     poke patterm's Unix socket directly. That socket connection
+//     carries no caller identity, so any sub-agent the agent spawns
+//     that way ends up as a stray top-level tab instead of a child
+//     under the spawning agent. Always go through the MCP tools.
+//   - The agent shells out to `claude` / `codex` / `opencode` to start
+//     a peer instead of calling `spawn_agent`. Those peers won't show
+//     up as sub-agents and won't be tied into the patterm lifecycle.
+//
+// Keep this short — clients vary in how much they surface to the LLM.
+const serverInstructions = "You are already running INSIDE patterm; the `patterm` MCP server is connected over the same stdio MCP transport you use for any other MCP server. Use the MCP tools you see in tools/list — do NOT (a) try to launch `patterm` or `patterm mcp-stdio` yourself, (b) poke the Unix socket through perl / nc / socat / curl, or (c) shell out to `claude` / `codex` / `opencode` to start a peer. Any of those bypasses caller-identity and the new agent will land as a stray top-level tab instead of a child under you. Start with `whoami` for your role and the full tool list, then `help('topics')` for orientation. `spawn_agent` is the only correct way to start a sub-agent; `spawn_process` is for non-LLM commands; `list_processes` / `get_process_output` inspect them; `send_input` / `send_message` drive them. Whatever you spawn is yours to `close_process` when done."
+
 // toolDescriptor is the shape returned by `tools/list`. inputSchema is
 // a JSON Schema object — we provide a minimal `{type: "object"}` schema
 // for each tool, which lets MCP clients accept arbitrary arguments and
@@ -88,7 +106,7 @@ func toolCatalog() []toolDescriptor {
 	return []toolDescriptor{
 		{
 			Name:        "spawn_agent",
-			Description: "Spawn a sub-agent from an agent preset and optionally seed it with initial instructions. Caller owns lifecycle: when the sub-agent's work is done (it reports back via send_message, or you no longer need it), call close_process on its process_id to free the pane and tear down the PTY. See help('lifecycle').",
+			Description: "Spawn a sub-agent from an agent preset and optionally seed it with initial instructions. This is the ONLY correct way to start a sub-agent under you — do not shell out to `claude` / `codex` / `opencode` and do not poke patterm's Unix socket via perl / nc / socat. Either bypasses caller identity and the new agent lands as a stray top-level tab instead of your child. Caller owns lifecycle: when the sub-agent's work is done (it reports back via send_message, or you no longer need it), call close_process on its process_id to free the pane and tear down the PTY. See help('spawning') and help('lifecycle').",
 			InputSchema: objectSchema(map[string]any{
 				"agent":              stringProp("Preset name (e.g. \"claude\", \"codex\")."),
 				"agent_instructions": stringProp("Initial prompt typed into the agent after it's ready."),
@@ -377,7 +395,8 @@ func (s *Server) handleProtocolMethod(callerID, method string, params json.RawMe
 			"capabilities": map[string]any{
 				"tools": map[string]any{"listChanged": false},
 			},
-			"serverInfo": serverInfo,
+			"serverInfo":   serverInfo,
+			"instructions": serverInstructions,
 		}
 		return result, true, 0, "", nil