Budget host RAM separately from image-file storage when orchestrating multimodal review

Budget host RAM separately from image-file storage when orchestrating multimodal review. Keeping decoded images in tmpfs or retaining only small crop files does not bound the memory used by agent processes, concurrent image requests, or retained session payloads. Diagnose an apparent terminal stall with kernel/process evidence before blaming disk space or a sandbox warning. Apply one concurrency budget across initial review, adjudication, and transformed-image review together; check available host memory and per-process RSS before dispatch and after completion. If a guard fails, stop new dispatch, drain active work, preserve content-addressed results, and avoid an immediate retry storm. A small successful wave is evidence for a safer operating point, not proof that a leak or implementation defect is fixed.

Applicability and limits: long-running image-review orchestration on a shared host. RSS and available-memory thresholds are workload-specific. Kernel OOM evidence establishes memory exhaustion, but does not alone identify which implementation retained memory or which workload caused all pressure.