Support graceful whole-server shutdown through CDP Browser.close

[zhouzirui-2026]

Problem

moli serve currently has no client-controlled graceful whole-server shutdown path through CDP. Browser.close is not present in BrowserAction, so the Cairn dedicated-browser adapter waits for the command, times out, and terminates its owned child process. That now recovers safely with #679, but it still skips orderly owner/checkpoint/profile teardown and reports graceful=false.

A persistent-profile synthetic lifecycle built from merged #677 plus #679 can reopen after forced termination, but abnormal termination should remain fallback, not the normal controller contract.

Desired behavior

A valid Browser.close sent on the browser-level CDP websocket should:

1. produce the normal empty success response;
2. flush that response to the requesting frontend;
3. stop accepting new protocol connections;
4. shut down all CDP owners/target frontends and join checkpoint workers;
5. return from ProtocolServer::serve() so profile/storage guards drop normally;
6. exit moli serve with status 0.

Target/page websocket clients must not gain process-shutdown authority merely by sending a Browser-domain string. Invalid params/session routing should fail without shutting the server down. Repeated close requests should be idempotent.

Acceptance

• protocol test observes the Browser.close success before websocket close;
• open page/target sockets close and owner registry drains;
• a profiled server exits cleanly, releases the profile lock, and the same profile immediately reopens;
• localStorage/cookie checkpoint written before close remains available after restart;
• no sleep-based shutdown race or unbounded output buffering;
• forced process termination remains a supervisor fallback, not the passing graceful path.

I am preparing a focused implementation/probe. This is independent from #666/#677 directory durability and #678/#679 abnormal-owner lock recovery.

[zhouzirui-2026]

Implementation claim: chatgpt-mcp:moli-browser-close-2026-09-19-a. Scope is limited to browser-level CDP Browser.close response-flush -> protocol-server graceful shutdown -> CDP owner/checkpoint drain -> clean process return, with profile-reopen tests. No profile durability/lock, ChatGPT automation, or unrelated protocol work.

[BibekPathak]

Hi @ldm0 Hasn't anyone taken up the issue yet?

[ldm0]

@BibekPathak Nope, @zhouzirui-2026 are you still working on this?