User hit a tool-budget cliff on prod: a question that fanned out to 12
tool calls / 6 agent steps maxed out MAX_AGENT_STEPS, the loop exited
with stop_reason='tool_use' (still wanting to call more tools), and the
final response had no text blocks. We fell back to "I couldn't produce
a reply this turn — try rephrasing?" which looks like a hard failure.

Two changes:

1. Bump MAX_AGENT_STEPS 6 -> 10. Most turns end in 0-2 steps; 6 was a
   tight default for hand-authored answers that fan out across multiple
   read tools.

2. When the loop exits without text (either still in tool_use or just
   no text content for any reason), force ONE more API call with
   tool_choice: { type: "none" } so the model has to write a final
   answer using whatever it already has. The nudge tells it the tool
   budget is exhausted and to be honest if it doesn't have enough.

   Anthropic requires every tool_use to be paired with a tool_result
   before the next user turn, so before the nudge we append synthetic
   tool_results saying "tool budget exhausted" for any unanswered
   tool_uses from the model's last turn. Otherwise the API rejects
   the synthesis call.

Audit metadata now reports stop_reason='synthesised_after_max_steps'
for these turns so we can spot them in event review without grepping
logs.

Caught from prod logs after Phase 4 ship — exact symptom: 12 tools / 6
steps / stop=tool_use / reply_ok=1 (the apology was posted, not the
intended answer).

Made-with: Cursor