inaccurate metrics

[kumanday]

The # of API calls doesn't seem correct. It's considerably lower than the number of steps. We need to audit & correct this, as well as other summary statistic metrics.

[kumanday]

This was an issue in non-dashboard mode, and appears to be working now in the default dashboard mode. Will keep monitoring, but it's probably OK.

Running the "review" task did surface heuristic approximation of tokens (chars/4), so we can look at getting actual numbers from LLM responses.