chore(deploy): switch the platform AI to DeepSeek deepseek-flash

Point ai.baseUrl at https://api.deepseek.com and ai.model at
deepseek-flash; the key moves in PROD_SECRET_VALUES (APP_AI__API_KEY),
which is why nothing here carries it.

Probed the endpoint before switching: /models lists deepseek-flash as
DeepSeek-V4.1-Flash (1M context, image input, effort levels low/high/max),
/chat/completions accepts our exact body shape including the aliyun-era
thinking_budget field, JSON mode returns the {content,highlights,risks}
shape we parse, and the chain of thought arrives in reasoning_content
exactly as the client already expects. Small structured tasks come back in
1.5-3.0s where the previous qwen3.8-flash took 42-77s, so the 600s request
timeout and the 2048 thinking budget stay as they are.

Only configuration changed - the image stays sha-6d1afbc; the rollout that
follows is the env change restarting the pods.