[FIX] Cron introspection still loops into repeated terminal calls after loop-guard hard stop (#624)

[Lexus2016]

Problem (observed in real usage)\nThe agent's own unattended evolution-introspection cron runs keep being killed by the loop-guard hard stop after the same run repeatedly calls terminal. In the last 7 days, terminal was the most consecutively repeated tool (up to 23 consecutive calls in a single stuck run, 27 sessions affected) and the most failed tool (19 failures). The hard stop fires, but the next cron run often re-enters the same pattern instead of changing strategy, so the cycle repeats daily.\n\n### Evidence (aggregated, anonymized)\n- Frequency: seen in 27 of 292 recent sessions; terminal max consecutive run = 23\n- Tool/area involved: terminal\n- Failure shape: repeated consecutive terminal commands, many failing, ignored advisory nudges until cron hard stop\n\n### Impact on real tasks\nEvolution-introspection (this job) and other cron tasks spend their entire iteration budget stuck in terminal loops, fail with a loop-guard error, and produce no useful output. Self-improvement stalls because introspection cannot complete.\n\n### Proposed direction\nAdd a per-run, per-tool back-off / circuit breaker for cron: after the loop guard hard-stops one cron run for repeated terminal failures, the next run of the SAME cron job should start with a directive to avoid the previously stuck tool for the first N turns and to use non-terminal alternatives (file reads, subagent delegation, repo_map). This is a cron-session-level learning layer on top of the per-turn loop guard.\n\n### Value\n- Impact: 0.95\n- Effort: 0.55\n- Priority Score: 1.48