Run the same-room acceptance choreography on a real mock-model bench #324

#198's P2 item "先通过真实房间入口的确定性测试" pulled ahead of the P1 adapter work: the acceptance shape currently runs only with scripted DSH session events (#220). Before freezing contracts for OpenCode, the same shape should pass through the real room entry — actual DSH sessions, the scripted mock model, real shell tool execution — still deterministic and token-free.

Acceptance:

• A smoke script drives a real dev/bench.sh bench: bootstrap → delegation → forced clarification (order-verified: the question strictly precedes any work) → requester-controlled task wait/answer (bookkeeping makes no model calls) → the member writes total.mjs and runs a narrow self-test via a real shell call, reporting the actual exit code → the requester asks the admin to verify → the admin writes its own independent check file and runs it (not restating the member's claim) → task return → requester acceptance, with the recorded answer intact.
• The scripted mock gains a 协作任务 group choreography with the two shell commands (write+selftest, independent verify) — no upstream behavior change.
• Two consecutive fresh-bench runs pass identically (determinism).
• npm run verify stays green; core/ and patches/ regenerated together (new patch topic 0016 for dev/mock-llm.mjs).

This is evidence for the "real room entry, deterministic model" tier only — not a real-model trial, human acceptance, or a performance comparison.