[codex] Filter engine config flags by command context

[xu16601526267]

Summary

This PR prevents resolver-only or LLM-only config knobs from being emitted as CLI flags for service wrappers that do not support them.

What changed

• adds engine/model-type aware config flag filtering
• passes model type through resolved config and runtime deploy requests
• keeps legacy flag behavior when no engine/model context is available
• preserves LLM flags for vLLM/SGLang/llama-style commands and qwen-asr vLLM wrappers
• adds coverage for runtime flag emission and K3S pod generation

Why

AIMA can resolve generic model config keys such as max_model_len, gpu_memory_utilization, and mem_fraction_static, but image/audio/TTS/OCR service wrappers often do not expose those flags. Emitting them caused one-click deploy failures such as server.py: error: unrecognized arguments: --max-model-len 8192.

Validation

• go test ./internal/knowledge -count=1
• go test ./internal/runtime -run TestConfigToFlags -count=1
• git diff --check

[xu16601526267]

Linux target validation passed on aitopatom-5ba7 (qujing@222.128.15.222:22222, Linux 6.11.0-1016-nvidia, aarch64) with temporary user-local go1.25.0 linux/arm64.

Validation run:

• git diff --check origin/develop...HEAD
• go test ./cmd/aima ./internal/knowledge ./internal/runtime -count=1
• go test ./... -count=1

Remote logs:

• related package logs: /home/qujing/aima-pr-ready-validation/logs-20260515T172839Z
• full package logs: /home/qujing/aima-pr-ready-validation/logs-full-20260515T172935Z