#753 run_goal: empty final answer (gemma returns no text after tool-call history replay)

Validated Phase 0 dogfood (runs d00b9072 auto, 8356369b HITL) complete successfully but agent_result='' . Cause: the no-tool/final branch replays the raw conversation (assistant tool_calls + tool results) to gemma4:31b-cloud with no tools and no system_prompt; the model returns empty content in that state. Fix: build a clean final-answer prompt for the reasoning model (e.g. system_prompt instructing it to answer the user's goal from the gathered observations, or summarize observations rather than replaying tool-call messages). Plumbing is correct; this is answer quality. Relates #749.