Force a final answer on terminal error

Parent: agent-core

If the agent loop ends in an error having produced zero assistant text (max iterations hit, hallucinated tool name, etc.), instead of surfacing an error it makes one more LLM call without tools, asking the model to answer from what it already has — in-persona, not a canned fallback (backend/internal/conversation/inference/agent_loop_budget.go: handleTerminalError → forceFinalAnswer, on its own 60 s budget over context.WithoutCancel so it survives the turn's expired deadline).

So a visitor always gets a graceful, in-character answer even when the tool loop breaks.