fix(acp): complete turns against agents that notify before session/new resolves

[Yevanchen]

Problem

Production issue YEF-1119: on an OpenClaw Computer, a Web Console message stalls forever in "Working on this computer…" — no first token, no reply, no visible failure. Authorization, catalog, and model binding were all verified healthy.

Root cause

The OpenClaw Gateway ACP bridge (openclaw acp 2026.7.1-2) writes session/update notifications for the newly created session before it writes the session/new response. Captured wire order (loopback gateway, verbatim):

1. → session/new (id 1)
2. ← notification session/update (session_info_update, new session id)
3. ← notification session/update (available_commands_update)
4. ← session/new result

AcpClientRequestHandler.#applyUpdate asserted the update against the native session id, which only registers after the session/new response resolves, so it threw ACP driver backend session is not initialized.; the update-inbox failure closed the connection and every fresh-session OpenClaw turn failed deterministically during ACP session setup — before any prompt could complete.

A second defect compounds it during teardown: waitForChildProcessExit waited on stdio close, but the transport keeps the child's stdout locked behind web-stream readers, so close can stay pending after the SIGKILLed process is gone. Stop then threw ACP agent process did not exit after force kill. for an already-dead child.

Fix

• acp-client-request-handler.ts: while no native session id is registered, hold incoming session/update notifications in a bounded queue instead of failing the transport. acp-driver-backend.ts applies them right after session setup registers the id; pre-registration updates for a foreign session are dropped with a warning.
• acp-agent-process.ts: observe process exit independently of stdio close (port of f3ccdb2 from the main line onto this production lineage), and re-check exit metadata after a timed-out wait.

Tests

• New tests/acp-early-session-update.test.ts: real-process stub agent reproducing the captured OpenClaw frame order; asserts the prompt completes, the deferred updates are applied, and a foreign-session early update is dropped without failing the turn. Red on the parent commit, green here.
• tests/acp-agent-process.test.ts: descendant-holds-stdio case updated to the exit-observation contract (stop resolves promptly; group SIGKILL still reaps the descendant).

Verification

• bun test on Linux (oven/bun:1.3.14-debian): 1086 pass / 0 fail.
• bun run tc, bun run lint, git diff --check: clean.
• End-to-end on Linux with the real openclaw gateway + openclaw acp + a loopback mock model through mosoo-computer's mosoo-driver-turn built at this commit: prompt completes in ~5s with stopReason=end_turn (regression wired into the mosoo-computer image build; companion PR).

Production consumer: mosoo-computer bumps MOSOO_AGENT_DRIVER_COMMIT to this branch head (companion PR references YEF-1119).