fix: stabilize NPC model decision fallback

[JOY (JOY)]

Summary

• extend the Unity agent-decision RPC timeout separately from normal Nakama calls
• pass an explicit DOS.AI timeout through the Nakama runtime HTTP call and persist model backoff for permanent NPC actors
• back off repeated invalid/empty/timeout/502 model decisions and return structured NPC intent rejections instead of throwing gameplay-rule exceptions

Local review

• Local code-review fallback: approved, no blockers found in the scoped diff
• Boundary check: Unity still calls Nakama only; DOS.AI key remains server-side; model output remains validated intent only

Verification

• npm run build (backend/nakama)
• npm test (backend/nakama)
• dotnet build Unity/SecondSpawn.AI.csproj --no-restore
• dotnet build Unity/SecondSpawn.Networking.csproj --no-restore
• git diff --check
• docker restart second-spawn-nakama and /v2/rpc/secondspawn_health smoke
• live DOS.AI smoke: one post-restart Nakama decision returned source=model in 1.89s; later direct api.dos.ai checks still showed upstream instability with 44.7s latency and 502 after 60s, so the PR hardens fallback/backoff but does not fix api.dos.ai/vLLM itself