Restored session permanently loses tool/subagent access when its persisted runtime binding is stale (acp-runtime-after-reap)

Bug

A session whose persisted agent runtime binding references a runtime that the current process never registers becomes permanently unable to execute tools or spawn subagents after being restored. Every tool call and subagent spawn fails with:

runtime acp:session_<id> does not exist in workspace <id>

The agent still converses (model calls need no runtime), but cannot act. Restarting turns or re-seeding does not help — only creating a brand-new session escapes.

Root cause

AgentRuntimeBindingService's restore hook (packages/agent-core-v2/src/agent/runtimeBinding/runtimeBindingService.ts) replays a persisted {workspaceId, runtimeId} binding back into state without checking that the runtime exists. RuntimeRegistry is purely in-memory per process. So when a session that was previously driven via kimi acp (binding persisted as acp:<sessionId> by bindSessionRuntime) is later resumed in a host that never registers ACP runtimes — the TUI, kap-server, or headless kimi -S <id> -p — the restored binding points at a runtime id that will never exist in that process:

• runtime-gated tools (Bash, Read, Write, …) stay deactivated because availability probes return false, and
• IAgentRuntimeService.acquire() throws runtime.not_found for every subagent spawn and any tool that does run.

Note that validating the binding at restore time is not a correct fix: in the healthy ACP flow, agent scope restore runs before bindSessionRuntime registers the session's ACP runtime (wireSession in packages/acp-server/src/server.ts), so restore-time acquisition would break the normal resume path.

Reproduction

1. Create a session over ACP (kimi acp + session/new), so the main agent's binding persists as acp:<sessionId>.
2. Resume the same session in a non-ACP host, e.g. headless: kimi -S <sessionId> -p "…".
3. Ask the agent to spawn a subagent → fails with runtime acp:session_<id> does not exist in workspace <id>; runtime-gated tools are missing from the toolset.

Proposed fix

Heal lazily at the point of use: in AgentRuntimeService.inspect()/acquire()/isAvailable() (packages/agent-core-v2/src/agent/runtimeBinding/agentRuntime.ts), catch RuntimeError with code runtime.not_found, rebind the agent to the always-present local runtime via IAgentRuntimeBindingService.set() (which persists the durable RuntimeSetBinding event), and retry once. Runtimes that are registered but unavailable (disconnected/draining) never trigger the heal; if the fallback is also missing, the original error is rethrown with the fallback failure attached as cause.

PR with the fix + unit tests follows, linked to this issue.

Start in packages/agent-core-v2/src/agent/runtimeBinding/agentRuntime.ts and trace inspect(), acquire(), and isAvailable() against AgentRuntimeBindingService.set(). Review the restore hook in packages/agent-core-v2/src/agent/runtimeBinding/runtimeBindingService.ts and ACP registration through wireSession in packages/acp-server/src/server.ts. Done means stale runtime.not_found bindings heal to local without healing registered-but-unavailable runtimes, and unit tests cover the retry and fallback error paths.