feat: RFC-011 -- local LLM backend selection (llama-cpp-python + onnxruntime-genai) (#104)

* docs: add RFC-011 for local LLM backend selection (llama-cpp-python + onnxruntime-genai)

Introduce a config-managed local_backend field under llm.provider: local
that lets users select between llama-cpp-python (current default) and
onnxruntime-genai as the in-process inference engine.

Key design decisions:
- Internal LocalBackend protocol keeps the change isolated to local_provider.py
- Zero caller impact — no change to generate() or any of its 7 call sites
- onnxruntime-genai is an optional dependency (pip install zettelforge[local-onnx])
- Execution provider (cpu, cuda, rocm, dml) passes through extra config
- Both backends share the same LocalProvider dispatcher and registry entry

Related: RFC-002 (Universal LLM Provider Interface)

* feat: implement RFC-011 — local LLM backend selection via config (llama-cpp-python + onnxruntime-genai)

Implements Phase 1 of RFC-011: adds local_backend config field under
llm.provider: local that selects between llama-cpp-python (default)
and onnxruntime-genai as the in-process inference engine.

Changes:
- Extract LlamaCppBackend class from LocalProvider (preserves existing behavior)
- Add OnnxGenAIBackend class with lazy SDK load and full generate() support
- Refactor LocalProvider to dispatch on backend parameter via _get_impl()
- Add local_backend field to LLMConfig dataclass + __repr__
- Add ZETTELFORGE_LLM_LOCAL_BACKEND env override in _apply_env()
- Forward local_backend through _provider_kwargs() in llm_client.py
- Add local-onnx and local-all optional extras to pyproject.toml
- Update config.default.yaml with local_backend documentation and examples
- Update test suite: +15 new tests for backend dispatching, all backends,
  and backward-compat _llm setter path
- Update RFC-011 with Decision and Implementation Status sections

RFC: docs/rfcs/RFC-011-local-llm-backend-config.md
PR: #104

* style: fix ruff format in local_provider.py

- Collapse single-line string expressions (ruff auto-format)
- Collapse single-line AttributeError messages (ruff auto-format)

CI fix: lint check (ruff format) was failing on 4 formatting issues.

* fix: backend-aware model/filename defaults in _provider_kwargs (RFC-011)

Bug: When local_backend: onnxruntime-genai was configured but llm.model
contained an Ollama tag (e.g., the default qwen3.5:9b), _provider_kwargs
unconditionally substituted DEFAULT_LLM_MODEL (a GGUF HuggingFace repo)
and DEFAULT_LLM_FILENAME (a .gguf file) regardless of which backend was
selected. The ONNX backend would receive a GGUF model reference.

Fix: _provider_kwargs now inspects local_backend before choosing fallback
defaults. When backend is onnxruntime-genai, it uses ONNX-appropriate
defaults (microsoft/Phi-3-mini-4k-instruct-onnx, .onnx filename).
When backend is llama-cpp-python (default), existing GGUF defaults apply.

Also adds _DEFAULT_ONNX_MODEL and _DEFAULT_ONNX_FILENAME constants to
llm_client.py for cross-referencing.

* style: fix ruff format in llm_client.py

Collapsed nested ternary expression that ruff format flagged.

Author: rolandpg

config.default.yaml

@@ -146,21 +146,34 @@ embedding:
 # extraction, and RAG synthesis.
 #
 # Provider selects the backend registered in zettelforge.llm_providers:
-#   local           — in-process llama-cpp-python (requires zettelforge[local])
+#   local           — in-process LLM inference (see local_backend below; requires zettelforge[local] or zettelforge[local-onnx])
 #   ollama          — Ollama HTTP API
 #   mock            — test-only canned responses
 #   openai_compat   — any /v1/chat/completions endpoint (Phase 2, upcoming)
 #   anthropic       — Anthropic native SDK (Phase 3, upcoming)
 #
+# local_backend selects the in-process inference engine when provider=local:
+#   llama-cpp-python     — GGUF models via llama-cpp-python (default; requires zettelforge[local])
+#   onnxruntime-genai    — ONNX models via ONNX Runtime GenAI (requires zettelforge[local-onnx])
+#
 # Examples:
 #   # Default — Ollama on localhost
 #   provider: ollama
 #   model: qwen3.5:9b
 #
 #   # In-process GGUF (fully offline)
 #   provider: local
+#   local_backend: llama-cpp-python
 #   model: Qwen/Qwen2.5-3B-Instruct-GGUF
 #
+#   # In-process ONNX (AMD GPU via ROCm)
+#   provider: local
+#   local_backend: onnxruntime-genai
+#   model: microsoft/Phi-3-mini-4k-instruct-onnx
+#   extra:
+#     filename: phi3-mini-4k-instruct-q4.onnx
+#     provider: rocm            # cpu, cuda, dml, rocm, openvino, coreml
+#
 # api_key supports ${ENV_VAR} references — never commit raw API keys.
 #   api_key: ${OPENAI_API_KEY}
 #
@@ -177,10 +190,11 @@ embedding:
 #   ZETTELFORGE_LLM_PROVIDER=ollama
 #   ZETTELFORGE_LLM_MODEL=qwen2.5:7b
 #   ZETTELFORGE_LLM_URL=http://gpu-box:11434
-#   ZETTELFORGE_LLM_API_KEY=sk-...
+#   ZETTELFORGE_LLM_API_KEY=***
 #   ZETTELFORGE_LLM_TIMEOUT=60
 #   ZETTELFORGE_LLM_MAX_RETRIES=2
 #   ZETTELFORGE_LLM_FALLBACK=ollama
+#   ZETTELFORGE_LLM_LOCAL_BACKEND=onnxruntime-genai   # only when provider=local
 #
 llm:
   provider: ollama
@@ -191,6 +205,7 @@ llm:
   timeout: 60.0
   max_retries: 2
   fallback: ""
+  local_backend: llama-cpp-python  # used when provider=local (RFC-011)
   extra: {}