Reference telegram brain iteration starvation fix

In scripts/forge_telegram_brain.py, the coordinator persona's MAX_ITERS was 4, but the CreatorTrack/email cutover (commits 243ed0e + 5d88385) made read paths multi-step (e.g. search_emails -> search_emails -> read_email -> synthesize = 5 turns). When the loop hit the cap mid-work, final_reply held an empty tool-call turn's text and return final_reply or "Done." shipped a bare "Done." even though the brain had already found the answer (e.g. read a Chase Travel reservation for a Mexico trip question). Same root cause produced empty replies on an earlier Austin FC question.

Fix applied 2026-07-03: (1) raised coordinator MAX_ITERS 4 -> 7 (general persona already ran 8 for the same reason; 180s wall-clock budget still bounds runaway loops), and (2) added a for/else synthesis pass that fires only on iteration exhaustion, making one no-tools brain call to force an answer from already-gathered tool_results with an explicit "never reply empty" instruction.

Why: silent "Done." replies erode trust in the coordinator bot and are hard to notice without reading brain trace logs. How to apply: if a Telegram bot persona later gets a deeper/multi-step tool catalog (more search/read hops before it can answer), check its MAX_ITERS against the new hop count, and confirm the synthesis-pass fallback still exists rather than a bare or "Done." return.

[auto-memory session 6b59fd6c-e0e6-4c96-a90d-d0dfe236981b, confidence 0.75, mode staged]