feat(llm): add Gemini CLI provider as primary; set qwen3.5:9b as default Ollama model

- Add GeminiCliProvider: shells out to `gemini -p` with --output-format json,
  injection-safe prompt passing, MCP server suppression via temp workdir,
  6-slot concurrency semaphore, 60s subprocess deadline
- Add --llm-provider, --llm-model, --llm-base-url CLI flags for per-call overrides
- Provider chain: Gemini CLI → OpenAI → Ollama → Anthropic
- Move LLM timing to dispatch layer (LLM: Xs on stderr)
- Default Ollama model: qwen3:8b → qwen3.5:9b (benchmark shows better schema extraction)
- Add noxa mcp subcommand
- Add docs/reports/llm-benchmark-2026-04-11.md (Gemini vs qwen3.5:4b vs qwen3.5:9b)
- Bump version 0.3.11 → 0.4.0

Co-authored-by: Claude <claude@anthropic.com>

crates/noxa-llm/src/testing.rs

@@ -4,8 +4,8 @@
 /// extract, chain, and other modules that need a fake LLM backend.
 #[cfg(test)]
 pub(crate) mod mock {
-    use std::sync::atomic::{AtomicUsize, Ordering};
     use std::sync::Arc;
+    use std::sync::atomic::{AtomicUsize, Ordering};
 
     use async_trait::async_trait;
 
@@ -50,7 +50,7 @@ pub(crate) mod mock {
     }
 
     /// A mock provider that returns responses from a sequence.
-    /// Call N → returns responses[N], wrapping at the end.
+    /// Call N → returns responses[N], clamping to the final response.
     /// Useful for testing first-failure / second-success retry paths.
     pub struct SequenceMockProvider {
         pub name: &'static str,
@@ -60,10 +60,11 @@ pub(crate) mod mock {
     }
 
     impl SequenceMockProvider {
-        pub fn new(
-            name: &'static str,
-            responses: Vec<Result<String, String>>,
-        ) -> Self {
+        pub fn new(name: &'static str, responses: Vec<Result<String, String>>) -> Self {
+            assert!(
+                !responses.is_empty(),
+                "SequenceMockProvider requires at least one response"
+            );
             Self {
                 name,
                 responses,