Complete Docker ecosystem integration and common LLM runtime flow

[abmmhasan]

Summary

Complete the LocalDevStack bottom-up Docker ecosystem integration and adapt local AI to the final mutually-exclusive provider architecture.

Local AI architecture

Exactly one provider is active per stack:

supported AMD XDNA2 NPU -> llm-fastflow -> infocyph/llm-fastflow:latest
NVIDIA GPU              -> llm-ollama  -> infocyph/llm-ollama:latest
AMD ROCm GPU             -> llm-ollama  -> infocyph/llm-ollama:amd-latest
CPU fallback             -> llm-ollama  -> infocyph/llm-ollama:latest

llm-fastflow and llm-ollama are mutually exclusive and never intentionally run together. The active service owns the common Docker alias llm:11434.

Provider-neutral consumers use:

LDS_AI_PROVIDER=llm
LDS_AI_URL=http://llm:11434
https://llm.localhost/v1
http://127.0.0.1:11434/v1

Provider-specific routes remain available for diagnostics/native operations.

Runtime/model behavior

• XDNA2 NPU detection wins automatically when supported.
• lds llm runtime auto|npu|nvidia|amd|cpu controls explicit runtime selection.
• FastFlow default model: qwen3.5:9b.
• Ollama default model: qwen3:14b.
• New setup leaves LDS_AI_MODEL blank so the active provider default applies.
• Graphify uses the common llm.localhost:11434/v1 route and validates the selected model through /v1/models.
• Local Graphify extraction uses ephemeral provider definitions and forces thinking off for strict structured output.
• LDS_AI_THINK / lds llm think <auto|on|off> provide one common thinking override across Ollama and FastFlow API-backed commands.
• LDS_LLM_ARCH has been removed; tracked Ollama uses :latest directly and the AMD runtime override selects :amd-latest directly.
• Nginx owns the loopback-only native port; provider containers remain internal.

Provider persistence

LLMModels         -> Ollama /root/.ollama
LLMFastFlowModels -> FastFlow /models

Neither provider receives a Docker socket or project/repository bind mount by default.

Other completed hardening

• dynamic LocalDevStack networks + safe legacy-network migration
• current moving image policy with the verified Elastic 9.5.4 exception
• selected PHP/Node Alpine runtime builds
• profile NONE/cancel/current-value semantics
• provider/tool stdin forwarding
• relative support-bundle output paths
• scoped cleanup and diagnostics
• current Tools/Runner/Nginx/Apache integration
• current published-image compatibility checks
• documentation and implementation-plan finalization

Validation

Final branch validation covers:

• static + ShellCheck
• CLI/environment/catalog/runtime/network/wrapper/security/QoL/docs contracts
• mutually-exclusive Ollama/FastFlow Compose matrices
• common OpenAI-compatible LLM fixture
• published llm-fastflow:latest and Ollama image references
• current Tools 0.25 common-llm consumer contract
• Windows bridge
• Sphinx documentation
• selected PHP/Node runtime builds

Plan

docs/plans/docker-ecosystem/07-localdevstack-integration-plan.md

The final AI provider override at the end of the plan supersedes the earlier Ollama-only planning text.