Summary

• Switch SubAgentActor.InvokeLlmAsync from GetResponseAsync to GetStreamingResponseAsync + updates.ToChatResponse(), matching the streaming path LlmSessionActor already uses via SessionLlmInvoker.StreamAsync. Non-streaming was dropping TextReasoningContent for providers that emit thinking blocks (e.g., Qwen), leaving the subagent's final assistant message with no TextContent and causing ExtractText to return the literal string "(no response)" while Complete reported success=True with 13 chars of output.
• Guard against future empty-response regressions: if ExtractText returns the empty-response marker, surface as Complete(false, ...) with a loud warning instead of fake success. Parent sessions used to receive (no response) as a tool result and hallucinate "subagent still processing" from it.
• Update SubAgents FakeChatClient to delegate GetStreamingResponseAsync through GetResponseAsync + ToChatResponseUpdates() so existing test cases exercise the new streaming path.

Why

Pre-fix daemon logs and live smoke testing both reproduced the symptom: research-assistant and code-analyst subagent spawns frequently returned 13-character outputs (exact length of "(no response)") while the actor reported success and iterations=0. From the parent LLM's perspective, spawn_agent appeared to return nothing, leading the frontline model to fabricate "the subagent is still processing" responses and to eventually stop reaching for the tool. This is the root cause of the "launch pad crash" and "0% success rate" symptoms reported in #264.

Test plan

• dotnet test src/Netclaw.Actors.Tests --filter "FullyQualifiedName~SubAgent" — 25/25 green
• Live smoke test (Phase 0.1): explicit-delegation prompt against Qwen3.5-27B-UD produced success=True, output=1065 chars, iterations=5, duration=49958ms with real cited findings, vs. the broken path's success=True, output=13 chars, iterations=0
• Live daemon deployment via local binary swap; sub-agent performed 10+ tool iterations of real web_search / web_fetch / file_read work before terminating naturally on a second successful run
• Phase 0.2 organic-delegation test already run: confirmed the streaming fix is orthogonal to the adoption gap — Qwen still does the research in-process instead of delegating. Adoption work tracked in Expand subagent delegation guidance #522 and the new sub-agent format work filed in feat(subagents): single-file markdown format + contextual prompt + project-scoped discovery #647.

Related follow-ups (not closed by this PR)

• Expand subagent delegation guidance #522 — expand sub-agent delegation guidance in AGENTS.md (adoption gap, orthogonal to this streaming fix)
• 🐛 Daemon silently dies on background-thread exceptions — no crash log, no notification #643 — daemon silently dies on background-thread exceptions (observability gap uncovered while smoke-testing this fix)
• feat(subagents): single-file markdown format + contextual prompt + project-scoped discovery #647 — single-file markdown format + contextual prompt + project scoping (format redesign)
• feat(providers): concurrent multi-model / multi-provider architecture for per-role and per-agent routing #648, feat(subagents): per-agent model selection in frontmatter #649 — multi-model provider architecture and per-agent model selection (follow-ons)

Closes #264