feat: OpenAI-compatible LLM gateway provider

[dpup]

Summary

Add an OpenAI decomposer to the LLM gateway alongside the existing Anthropic one. The gateway architecture already supports multiple providers — internal/gateway/anthropic/ is self-contained with a clean decomposer pattern. This is implementation work, not design work.

Scope

• New package: internal/gateway/openai/
• Decompose OpenAI chat completion requests into Keep calls:
  • llm.request — model, token estimate, system message summary
  • llm.tool_result — tool role messages (function call results)
  • llm.tool_use — assistant tool_calls in responses
  • llm.response — completion summary (finish_reason, tool_call count)
• Support both regular and streaming (stream: true with SSE) responses
• Mutation patching for redact actions (same pattern as Anthropic)
• Gateway config: provider: openai

OpenAI-specific considerations

• Chat completions use messages[] with role-based content (system, user, assistant, tool)
• Tool calls are in tool_calls[] on assistant messages, not content blocks
• Streaming uses data: [DONE] sentinel vs Anthropic's event types
• Function calling has both legacy function_call and modern tool_calls — support modern only

Out of scope

• OpenAI Responses API (different shape entirely — future issue if needed)
• Assistants API / threads

References

• Existing Anthropic decomposer: internal/gateway/anthropic/
• Gateway proxy: internal/gateway/proxy.go
• PRD R-3: "OpenAI-compatible as a fast follow"
• gateway/config/config.go:54 — placeholder comment