Thinking summary moved into the composer; migrated service tier / verbosity dropped

What changed

OpenAI models now return a thinking summary by default, and its verbosity (Auto / Concise / Detailed) is picked per conversation from the speed control next to the send button — the same popover that holds the reasoning effort slider. The row only appears for hosts that accept the parameter (OpenAI, Azure OpenAI, Codex).

Three provider-level settings carried over from v1 are gone: reasoning summary, service tier, and verbosity.

Why this matters to the user

Before this change, OpenAI reasoning models (gpt-5, o-series) showed no thinking content at all in Cherry, because the summary parameter was never sent. After upgrading, thinking appears without any configuration.

Service tier and verbosity were only reachable by users who migrated from v1 with those values already set — v2 never offered a way to view or change them. Those values now stop being applied. Requests otherwise behave the same; low-latency inference is still available through the Fast (⚡) button in the same popover.

What the user should do

Nothing — automatic. To change the summary verbosity, open the speed control in the composer and pick Concise or Detailed.

Notes for release manager

Third-party OpenAI-compatible hosts (Volcengine Ark, Alibaba DashScope, DeepSeek, …) are unaffected: they emit thinking summaries on their own and reject the parameter, so the control is not shown for them.