In-process desktop (Tauri) core runtime still uses the default 2 MiB worker stack — subagent delegation can crash it #3159

Follow-up to #3155.

#3155 fixes the standalone openhuman-core run JSON-RPC server by setting thread_stack_size(16 MiB) on its tokio runtime (src/core/cli.rs), so a sub-agent delegation no longer overflows the default 2 MiB worker stack and aborts the process (SIGABRT, thread 'tokio-rt-worker' has overflowed its stack).

The in-process desktop core runs on a different runtime — the Tauri host's tokio runtime (app/src-tauri), spawned via core_process::CoreProcessHandle — which is not covered by that change and still uses the default 2 MiB worker stack. The same crash is therefore still reachable in the shipped desktop app whenever the orchestrator delegates to a sub-agent with a large tool surface / prompt.

Repro / mechanism

Identical to #3155: orchestrator runs a research-style tool → subagent_runner dispatches a sub-agent at spawn_depth=1 → the nested agent-turn future overflows the worker stack. Reproduced for the CLI path by driving chat-harness-subagent.spec.ts against a real openhuman-core (core aborts with Abort trap: 6, fatal runtime error: stack overflow).

Fix

Apply the same thread_stack_size(16 MiB) to the tokio runtime that the embedded desktop core runs on (and ideally any other multi-thread runtime in src/core/* that can run an agent turn — agent_cli, call, etc.). Consider a shared constant so the value stays in sync.