fix(agents): correctly recognize prefixed GPT-5.6 models in opencode and pi

[Filip-Stastny-Philips]

Summary

Fixes model classification and Responses API routing for GPT-5.6 that include provider prefixes (e.g., openai.gpt-5.6-luna) in OpenCode and Pi agent plugins. Previously they were not recognized and fell into the default 120k input / 4k output limits, which is practically unusable.

Changes

• Updated Responses API pattern matching in OpenCode and Pi to recognize prefixed and hyphenated GPT-5.6 models
• Fixed OpenCode model family detection and token limit heuristics for prefixed GPT-5.6 variants
• Enabled reasoning mode and correct context window limits for prefixed GPT-5.6 models in Pi
• Fixed MCP server stdio transport import path in bridge module

Impact

• Before: Prefixed models like openai.gpt-5.6-luna fell back to generic chat completions or default token limits and disabled reasoning.
• After: Prefixed GPT-5.6 models correctly route to /v1/responses, receive the 1.05M context window, and enable reasoning effort.

Checklist

• Self-reviewed
• Manual testing performed
• No breaking changes (or clearly documented)