feat: real LLM call via wasi-http (Ollama adapter) — v0.2.0

The orchestrator now calls a configured Ollama-compatible LLM
endpoint over wasi-http and emits the LLM-generated plan back
to the hub. When llm_endpoint is empty, the v0.1.0 deterministic
stub is preserved as fallback (useful for tests and air-gapped
environments).

- module.yaml: bump to 0.2.0; new inputs llm_endpoint,
  llm_model, llm_api_key; net: permissions for OpenAI,
  Anthropic, localhost, 127.0.0.1 (Ollama defaults)
- src/llm.rs: LlmClient trait, OllamaParams, build_user_prompt,
  build_ollama_body, extract_ollama_content, generate_plan;
  9 unit tests for prompt-building and response-parsing
- src/plan.rs: tagged enum with Explain/SaveFlow/RunFlow/
  InstallModule variants; parse_and_validate; build_stub_plan
  (renamed from build_plan); validate(); 6 unit tests
- src/lib.rs: WakiClient implementing LlmClient over the waki
  WASI-0.2 HTTP crate (wasm32-only); empty endpoint → stub
- Cargo.toml: 0.2.0; thiserror dep; waki dep gated on wasm32

17 host-side tests pass. WASM bundle is 386KB (was 74KB; the
delta is wasi-http via waki, expected). The deny path is
exercised by the platform-side integration test in
fai_platform/crates/fai_hub/tests/orchestrator_llm_path.rs.

Signed-off-by: flemming-it <sf@flemming.it>

Cargo.toml

@@ -5,7 +5,7 @@
 
 [package]
 name = "orchestrator_llm"
-version = "0.1.0"
+version = "0.2.0"
 edition = "2024"
 authors = ["Dr. Stefan Flemming <platform@flemming.ai>"]
 license = "Apache-2.0"
@@ -20,6 +20,11 @@ crate-type = ["cdylib", "rlib"]
 wit-bindgen = "0.36"
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
+thiserror = "2"
+
+# WASI-0.2 HTTP client. Only compiled into the wasm32-wasip2 build.
+[target.'cfg(target_arch = "wasm32")'.dependencies]
+waki = "0.5"
 
 [dev-dependencies]
 serde_yaml = "0.9"