#788 Agent loop can 'complete' at turn 0 with zero tool calls on a structured/constrained goal

A run_goal_deferred with a 'Step 1: delegate... do NOT research yourself; Step 2: do A/B/C' goal completed at turn 0 having executed NOTHING -- the model replied with prose ('I cannot, the steps were not executed, I'm instructed not to call tools now') instead of calling the delegate tool, and the while_loop treated the no-tool-call text as agent_done -> ended. An action-oriented agent should never finish having made zero tool calls. FIX: (a) loop guard -- if the goal requires actions and no tool was called, re-prompt/continue rather than mark done; (b) softer multi-step prompting (over-constraining 'do not do X' made the model describe instead of act). Repro: tenant demo run 19a168fb. Contrast: a natural unattended goal (run 53018c1b) executed all 3 actions correctly.