fix(skills): align llm-provider stream() signature across platforms

[xiaolai]

Automated audit: This PR was generated by NLPM, a natural language programming linter, running via claude-code-action. Please evaluate the diff on its merits.

Bug

The .agents/ and .cursor/ copies of skills/llm-provider/SKILL.md document incompatible LLMProvider.stream() signatures that contradict the actual interface in src/llm/types.ts:

Platform	Documented stream() signature
.claude/	stream(options: LLMStreamOptions, callbacks: LLMStreamCallbacks): Promise<void> ✅
.agents/	async *stream(request: LLMRequest): AsyncGenerator<string> ❌
.cursor/	async *stream(messages: ChatMessage[], params: ChatParams): AsyncIterable<StreamChunk> ❌

The actual interface (from src/llm/types.ts):

export interface LLMProvider {
  call(options: LLMCallOptions): Promise<string>;
  stream(options: LLMStreamOptions, callbacks: LLMStreamCallbacks): Promise<void>;
  listModels?(): Promise<string[]>;
}

Why it matters

An agent following the .agents/ or .cursor/ skill will generate a provider class with a generator-based stream() method that doesn't match the LLMProvider interface. This produces a TypeScript compile error (does not implement LLMProvider) before any code runs, breaking provider development for anyone using those platform skills.

Fix

Updated .agents/ and .cursor/ to document the correct callback-based signature, matching .claude/ and the actual src/llm/types.ts definition. Import types and constructor pattern also corrected in the .cursor/ version which had diverged more significantly (was using ChatMessage[], ChatParams, ChatError, StreamChunk — none of which exist in this codebase).