LONG THREAD / REAL SERVER HTTP MEASUREMENT
BEFORE  93f8563a95e08279cd770beb7aa6ff9791791049
Linux x86_64, Node v26.13.1, loopback HTTP, real T3 server, fresh synthetic SQLite database, same VM
Fresh replay history: 20 turns x 200 tools x 40,000 bytes
Prompts: agent/mcp; one inline tool every 20 turns; deferWork=1
3 warmups; 40 open + older reads; 15 x 6 concurrent opens
Five rounds, same VM, milliseconds

Round            Open        Older     Combined     6-open p95
1               18.94      1117.27      1136.62        82.15
2               20.26      1168.87      1190.08        84.86
3               17.54      1098.95      1115.84        78.36
4               18.02      1153.25      1171.65        85.56
5               16.48      1062.08      1077.84        68.92

Median of rounds / range across round statistics:
Open                 18.02 ms   [16.48, 20.26]
Older page         1117.27 ms   [1062.08, 1168.87]
Open + older       1136.62 ms   [1077.84, 1190.08]
Six opens p95        82.15 ms   [68.92, 85.56]

Returned rows: initial 75, older 75; deferred groups 0/19
Known inline items retained: 0/1 (initial/older); bodies load separately

Synthetic data only. Timings include loopback HTTP and JSON.
Not a browser/render measurement or the historical 6.3 GB DB.
