Release v3.2.0

Features

• Support OpenAI's Responses API for GPT-5.6 models
• Add LFM2.5 1.2B Instruct to the on-device model catalog

Fixes

• Cap stored message size so a large attachment can't brick the app
• Fix the Debian sandbox install failing on the base package step
• Send the OpenCode session header so OpenCode Zen no longer rejects requests

Improvements

• Let on-device models declare their own tool support and sampling defaults, so models without a tool template no longer invent answers
• Refresh the on-device catalog: updated gemma-4-12b-it build with vision, audio, and multi-token prediction
• Refresh model catalog Elo scores
• Upgrade dependencies