feat: implement LLM integration for team composition engine

Resolves WHOOSH-LLM-002: Replace stubbed LLM functions with full Ollama API integration

## New Features
- Full Ollama API integration with automatic endpoint discovery
- LLM-powered task classification using configurable models
- LLM-powered skill requirement analysis
- Graceful fallback to heuristics on LLM failures
- Feature flag support for LLM vs heuristic execution
- Performance optimization with smaller, faster models (llama3.2:latest)

## Implementation Details
- Created OllamaClient with connection pooling and timeout handling
- Structured prompt engineering for consistent JSON responses
- Robust error handling with automatic failover to heuristics
- Comprehensive integration tests validating functionality
- Support for multiple Ollama endpoints with health checking

## Performance & Reliability
- Timeout configuration prevents hanging requests
- Fallback mechanism ensures system reliability
- Uses 3.2B parameter model for balance of speed vs accuracy
- Graceful degradation when LLM services unavailable

## Files Added
- internal/composer/ollama.go: Core Ollama API integration
- internal/composer/llm_test.go: Comprehensive integration tests

## Files Modified
- internal/composer/service.go: Implemented LLM functions
- internal/composer/models.go: Updated config for performance

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>

internal/composer/models.go

@@ -215,14 +215,14 @@ type FeatureFlags struct {
 // DefaultComposerConfig returns sensible defaults for MVP
 func DefaultComposerConfig() *ComposerConfig {
 	return &ComposerConfig{
-		ClassificationModel: "llama3.1:8b",
-		SkillAnalysisModel:  "llama3.1:8b", 
-		MatchingModel:       "llama3.1:8b",
+		ClassificationModel: "llama3.2:latest",  // Smaller 3.2B model for faster response
+		SkillAnalysisModel:  "llama3.2:latest",  // Smaller 3.2B model for faster response
+		MatchingModel:       "llama3.2:latest",  // Smaller 3.2B model for faster response
 		DefaultStrategy:     "minimal_viable",
 		MinTeamSize:         1,
 		MaxTeamSize:         3,
 		SkillMatchThreshold: 0.6,
-		AnalysisTimeoutSecs: 60,
+		AnalysisTimeoutSecs: 30,  // Reduced timeout for faster failover
 		EnableCaching:       true,
 		CacheTTLMins:       30,
 		FeatureFlags: DefaultFeatureFlags(),