[CHORE] Choose an outlier-resistant statistic for benchmark charts

[lukekarrys]

Context

Registry charts publish the mean of a small number of runs (three for registry variations). One outlier can dominate the daily point: a single 294-second run produced a roughly 100-second mean during the incident. Standard deviation and median are already present in the raw data.

This is a methodology decision rather than a one-line display change. Options include:

• plot median;
• retain mean and add error bands/confidence information;
• separate run 0/cold behavior from steady-state runs.

This should compose with the partial-failure work so timeout survivors are not treated as complete samples.

Acceptance criteria

• The canonical statistic for benchmark charts is chosen and documented.
• The UI exposes enough variability/sample information to interpret small-run results.
• Cold/run-0 behavior is either intentionally included or intentionally separated.
• Processing, latest results, history, summaries, and leaderboard calculations use compatible semantics.
• Tests cover a sample containing one extreme outlier.

Source

Performance investigation: item 7 — mean-of-three amplifies outliers