ACP runs die within seconds on dev static

Found by the Task 7281 acceptance run on dev (CP 1.2.1940; static node ts2-node6, moved into gmind; project mr-6382 = ruzhukov/jaah-6382-backend). With a provider's ACP switch ON, every run on the ACP host ends within seconds, and no ACP event is ever ingested: - Claude Code, API key (claude-agent-acp): dev tasks 5089, 5098, 5104, 5106. Each ends "Runner terminated — session vanished" 1–12 s after start. On 5104 the agent did raise the question card ("Question raised" 22:32:57Z); the session then vanished at 22:32:58Z, the question was dropped, and the task went back to Queued. 5106, which asks nothing, also vanished after 6 s. - DeepSeek (claude-agent-acp): 5086 and 5097, the same "session vanished". - Pi (pi-acp): 5096. "broken session: unknown — Internal error: Cannot call write after a stream was destroyed" on every attempt, then system_failed. - For every one of these sessions, GET /api/tasks/<id>/sessions/<sid>/events returns `runtime: legacy` with 0 events. Control (AC-2): with claude:api_key switched OFF, task 5102 on the same node, account and repo ran legacy and finished `done` with merged PR ruzhukov/jaah-6382-backend#3 in 73 s. So the node, account, key and repo are healthy, and the fault is in the ACP launch, the host lifetime or the ingest. Find and fix why a jaah-acp-host run on a dev static node dies early, why the CP drops a question raised from it, and why no events or `runtime: acp` reach the CP. Use node logs and the replay harness. The console shows no node logs (see draft 7461 item 1). Out of scope: Codex/Gemini (no reachable account on dev, see draft 7458); MicroVM (draft 7459); deleting legacy paths. Done means: - On dev, a claude:api_key, a deepseek and a pi coder task each show live Agent activity events, a question card whose typed answer reaches the agent, and end `done` with a merged PR on mr-6382. - Their sessions show `runtime: acp` in Audit › Sessions, and the session_start event carries `cli_versions` equal to packaging/platform-pins.toml (AC-8). Required checks: ./dc check full

Found both causes (CP 422'd every event batch; a claude-agent-acp engine's own session read live runs DEAD). Same fix landed in #6036 (7467). Merged #6068 for the SDK v2 runtime gap. Pi cause reproduced, draft 7526. Dev re-run left to 7467. Full gate red only from develop (7531).