v0.28.46

Changed

• This release drops the unused completion_kit_mcp_sessions table. Run bin/rails completion_kit:install:migrations && bin/rails db:migrate. Nothing has read or written the table since 0.28.45, so deploy 0.28.45 to every instance first. The migration uses if_exists, and its rollback rebuilds the table.
• Engine copy no longer uses em dashes. Model groups, prompt versions, new auto run names and page titles now use a middle dot, such as "OpenAI · GPT-5" and "Support Reply Generator · v1 #3". Run names saved before this release keep their old form.

Fixed

• A prompt or run form rejected for its model keeps the choice and shows the error beside the select. The rejected model stays selected, marked "(unavailable)", and the select keeps its full width with the same red border as other invalid fields.
• Saving a run keeps a judge that has no model row. The run form used to show such a judge as "None", so a save cleared it. A save that forks a run with results also keeps a judge you did not change, and a run whose judge is hidden because no judge models are listed can be saved.
• A rerun is refused before anything is created when the generation API key is missing. It used to leave a failed run on every click. Run#rerun! raises CompletionKit::UnknownModelError when a model does not resolve and CompletionKit::ConfigurationError when the client is not configured.
• Unticking every metric on the run edit form now removes them. The form used to send nothing, so the run kept its metrics.
• The metric group form shows its errors at the top, and its breadcrumb and delete prompt keep the saved name after a failed save.
• Ticking "Score existing outputs" no longer returns a 500 in a host that raises on unpermitted parameters. The hint now says existing outputs will be judged.
• Renaming a run with results no longer forks it when the run has no output column set.
• A failed row whose provider is unknown shows its message without a stray colon.