(MOT-4277) feat(harness): expand benchmark execution insights

.github/benchmark-site/README.md

@@ -28,5 +28,24 @@ The execution index retains 100 workflow attempts. The latest 30 also retain the
 complete structured `results.json` content, including prompts, transcripts,
 session ids, gates, criteria, failures, retries, usage, and traces. The UI loads
 those reports only on the detail page and renders transcript-heavy sections only
-when expanded. Diagnostic logs, stack files, and credentials remain in
-access-controlled Actions artifacts.
+when expanded. Each run opens its transcript in a read-only dialog patterned
+after the Harness chat, with message and error filters, paired function calls
+and results, and recovered errors expanded by default. Diagnostic logs, stack
+files, and credentials remain in access-controlled Actions artifacts.
+
+Each full execution summary also carries compact per-scenario averages for
+tokens, wall time, cost, function calls, function-call errors, sessions, and
+turns. Tokens mean input plus output; cache-read tokens are already represented
+in input usage and are not added again. The execution table also exposes exact
+total tokens and function calls for every retained full report.
+
+Efficiency is the primary overview. Its cards show the current operational suite
+totals, while deltas use only successful scenarios with the same subject,
+scenario id, and behavioral contract fingerprint. New and changed scenarios
+collect five comparable executions before receiving a trend verdict. Removed
+scenarios remain visible as historical rows and never count as an efficiency
+gain. Contract changes start a new baseline instead of joining incompatible data.
+Select any scenario in the efficiency table to compare that scenario execution
+by execution. The modal switches between cost, tokens, duration, function calls,
+and function errors, marks contract boundaries, and links each point to its full
+execution details.