feat: M7 proper LLM integration with dual providers, caching, and parallel batching

- Add LLM client abstraction (src/llm-client.ts) with factory pattern
- Add OpenAI-compatible external client (src/external-llm-client.ts)
- Add Kimi.com client using Anthropic-based API (src/kimi-llm-client.ts)
- Add Pi native LLM stub (src/pi-llm-client.ts) for future ExtensionAPI wiring
- Add SHA-256 disk cache at ~/.cache/pi-project-map/ (src/llm-cache.ts)
- Add parallel batching with p-limit, retry + exponential backoff (src/llm-batch.ts)
- Rewrite llm-extract.ts to use real LLM calls with structured prompts
  - File-level: PURPOSE, DEPS, CONCEPTS
  - Package-level: ROLE, ARCH
  - Context truncation, 50KB skip, cache before LLM call
- Wire CLI with --llm-provider, --llm-model, --llm-base-url flags
- Update config.ts with llmProvider, llmBaseUrl fields
- Update init.ts and patch.ts to accept optional LLMClient
- Add sample project fixture for manual testing
- Add tests: llm-cache (3), llm-batch (5), llm-integration (8 with real Kimi API),
  pi-extension (14 mocked)
- All 56 tests pass

src/config.ts

@@ -4,7 +4,9 @@ import { join } from "path";
 export interface SkillConfig {
 	ignorePatterns: string[];
 	smallPackageThreshold: number;
+	llmProvider: "openai" | "kimi" | "pi";
 	llmModel: string;
+	llmBaseUrl?: string;
 	contextBudget: number;
 	autoInjectPrompt: boolean;
 }
@@ -32,6 +34,7 @@ export const DEFAULT_CONFIG: SkillConfig = {
 		".prettiercache",
 	],
 	smallPackageThreshold: 10,
+	llmProvider: "openai",
 	llmModel: "gpt-4o-mini",
 	contextBudget: 4000,
 	autoInjectPrompt: true,