Benchmark harness that records only p50/p95/min/max per query group cannot answer per-request questions later

A read benchmark grouped 20 seeded single-signal queries into one group row with n, p50, p95, min, max and a total. Afterwards a reviewer asked for one outlier request time and the p50 without that request; neither could be computed, only bounded (the 19-sample median lies between the two middle order statistics, i.e. within [min, p95] of the 20). The fix is cheap: emit a per-request row (query id, key, items returned, wall, cpu) next to the group summary, or at least the full sorted list of times. Also record which request produced the max.

Related: when looking up the queried names in a second copy of the data to attribute counts, use the exact file the benchmark read; a sibling file with a slightly different declaration set made one name look missing.