Fix/tool calls default metric

[jay-m-dev]

This pull request updates the handling of built-in runtime metrics, especially contextual metrics like tool calls and critic gates, to ensure only metrics that can actually be produced by the current canvas are pre-selected or enabled by default. It also removes "tool_calls" from the list of recommended metrics for new experiments, and updates related logic and tests accordingly. Additionally, it improves how node run statuses are sourced and displayed in the protocol canvas, ensuring consistency between test and real runs.

Metrics selection and availability improvements:

• Only metrics that the canvas can actually produce are pre-selected by default in the metrics editor, preventing users from saving metrics that can never report data. Contextual metrics (like tool_calls and critic_approvals) are now disabled and not pre-checked if the canvas lacks the necessary nodes. [1] [2] [3]
• The checkbox for an unavailable metric is now only disabled if it is not already selected, allowing users to uncheck metrics that became unavailable after being selected.

Recommended metrics changes:

• "tool_calls" is removed from the list of recommended runtime metrics, so new experiments will not include it by default unless the canvas supports it. Corresponding test expectations and catalog flags are updated. [1] [2] [3] [4] [5] [6]

Protocol canvas node run status consistency:

• The protocol canvas now consistently uses the latest available node run data (from either a real run or a test run) for node badges and inspectors, ensuring UI consistency and correct display after page reloads or test runs. [1] [2] [3] [4] [5]

Metrics plan synchronization:

• When saving metric selections, the logic now ensures that deselected metrics are fully removed from both the UI and the measurement plan, even if the design's metric list has changed, preventing potential errors on save.