Measure AI latency, token use, repairs, and edit isolation #34

Goal

Create a trustworthy baseline before changing providers or prompt context.

Scope

• Record time to first token and total time per Input, Command, and Advisor request.
• Record input/output token counts when a provider supplies them; otherwise retain a clearly marked size estimate.
• Record provider, model, lane, cancellation, repair attempts, final outcome, and affected block count.
• Extend eval reporting with edit success and unrelated-block modification checks.
• Never log prompt contents, API keys, or user document text.

Acceptance criteria

• The same repeatable scenarios can be compared before and after later roadmap work.
• Missing provider usage data is represented honestly, not as zero.
• Instrumentation has negligible user-visible overhead.
• Development/eval output explains the metrics without exposing note contents.

This is an enabling issue; it does not change provider selection or prompt construction.