Dream inference template conformance and reflection deepening

[sheawinkler]

Summary

• enforce final-content handling for OpenAI-compatible/Ollama inference responses and expose template conformance in runtime policy
• add Dream Mode reflection plus bounded deepening when output misses the sigma target
• add MLX Qwen final-content template, generic conformance fixtures, and runtime smoke scripts
• deep-clone source fanout request payloads to avoid continuation concurrency races

Verification

• GOCACHE=/Volumes/wd_black/codex_go_cache GOMODCACHE=/Volumes/wd_black/codex_gomod_cache go test -count=1 ./... (services/gateway-go)
• bash -n inference/runtime scripts; JSON template validation; contextlattice-dream py_compile