#768 Agent robustness: switch model to minimax-m3, add token-budget + gmail body ceiling, fail-fast reaper

Model A/B (gemma4:31b vs glm-5.2 vs minimax-m3) on an unanswerable question: gemma over-reached, ballooned context to 73.6k tokens and FAILED; glm over-reached and confabulated a scraped Google Ads ID as the account ID; minimax-m3 stopped at turn 0 with a clean admission. Over-reach also caused resource pathology (huge contexts, claim-expiry stuck runs, reaper thrash). FIXES: (a) default agent model gemma4:31b-cloud -> minimax-m3:cloud across enterprise/tools/agent.py defaults, api/services/reactions.py, api/dsl_templates/agent_run_goal.py, highway-mcp server.py (run_goal/scheduled/deferred). Validated minimax: restraint turn-0 admits ignorance, AND tool-use works (clean shell.run date -u, turn 1). (b) context ceiling: run_goal token_budget default None -> 60000; gmail get_message body capped to 4000 chars. (c) reaper fail_single_run sets LOCAL lock_timeout=5s + statement_timeout=20s so a locked stuck-run row is retried next cycle instead of tying up a worker for 2min and aborting mid-transaction (the prior '1 errors').