fix(engine): use user role for sub-agent completion runtime message

[h3c-hexin]

Problem

subagent_completion_runtime_message builds the sub-agent completion hand-off that gets appended into the parent turn with role: "system". That message is injected mid-conversation (not at the front).

Some OpenAI-compatible backends apply a strict chat template that requires any system message to be messages[0]. The clearest case is vLLM serving Qwen3 — its chat template raises:

System message must be at the beginning

vLLM returns this as a 400 BadRequest, which fails the request that carries the sub-agent result back to the parent. The whole sub-agent hand-off breaks on these backends.

Fix

Switch the role of this single internal runtime message from system to user.

The "this is an internal runtime event, not user input" framing is already carried by the visibility="internal" tag and the surrounding instruction text, so the role itself carries no semantic weight to the model here — it only determines chat-template placement. Using user keeps the message valid mid-conversation on strict templates while reading identically to lenient ones (Anthropic / OpenAI proper).

Scope

• One-line behavior change in subagent_completion_runtime_message, plus a rationale comment.
• Updates the colocated unit test (subagent_completion_handoff_is_internal_user_message) to pin the new role.
• No trust-boundary surface touched (no auth / sandbox / publishing / prompts).

Test

cargo test -p codewhale-tui --bins subagent_completion_handoff

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request changes the role of subagent completion runtime messages from 'system' to 'user' in crates/tui/src/core/engine/turn_loop.rs. This modification addresses compatibility issues with OpenAI-compatible backends, such as vLLM serving Qwen3, which require system messages to be positioned at the start of a conversation. Corresponding unit tests were updated to reflect this change. I have no feedback to provide.