Bug: Ollama 4GB memory limit insufficient for 7B models

[mrveiss]

Problem

docker-compose.yml limits Ollama to 4GB RAM. The default model mistral:7b-instruct requires ~4.5-5.5GB (4GB weights + runtime overhead). This will cause OOM kills under any concurrent load.

Discovered During

Code review of PR #1930 (#1897)

Impact

Severity: high — default Docker deployment will OOM on inference

Fix

Increase to 8G minimum, or make configurable via env var. Dev override (PR #1935) should also relax this.

[mrveiss]

Fixed in PR #1930. Ollama memory limit increased from 4G to 8G.