REAL SERVER PERFORMANCE COMPARISON
Linux x86_64; AMD EPYC 9455; Node v26.12.1
Base 2dfc4a9c983ef732083bc3d2580a75cf7a677db0
Head b9f5a450c5f5ad2c141985cac2601f7598ad3939
Fresh synthetic state on each side; 100 cycles of 400 tool calls + 7 app-owned delegates.
Nine rounds per parent; three quick tasks/results/status reads then three long tasks/cancellations.
All rounds retained. Median = median of round medians. Spread = max minus min round median.
Stable improvement requires gain > max(base spread, head spread), as requested.
Store reads are existing readCanonicalProjection spans (see analyze.py); operation times include other work.
Terminal finalization includes normal completion and cancellation; parent status-sync excludes its nested terminal finalization.
No heap claims; no isolated database-copy benchmark.

Task start (whole operation) / large: 232.174 [190.041-324.242] -> 150.289 [119.857-186.876] ms; gain 81.885; max spread 134.201; stable=False
Task start (whole operation) / tiny: 175.193 [146.469-206.351] -> 135.815 [123.788-158.120] ms; gain 39.377; max spread 59.882; stable=False

Task finish (whole operation) / large: 1049.135 [894.142-1258.283] -> 163.083 [133.941-205.739] ms; gain 886.052; max spread 364.141; stable=True
Task finish (whole operation) / tiny: 147.372 [96.278-209.107] -> 111.282 [102.885-140.431] ms; gain 36.090; max spread 112.829; stable=False

Parent status-sync (exclusive operation) / large: 90.170 [78.957-116.900] -> 33.467 [27.225-43.624] ms; gain 56.703; max spread 37.943; stable=True
Parent status-sync (exclusive operation) / tiny: 32.630 [21.739-40.707] -> 26.187 [17.686-30.150] ms; gain 6.444; max spread 18.968; stable=False

Task cancel (whole operation) / large: 1417.633 [1131.106-1553.981] -> 338.199 [281.187-436.789] ms; gain 1079.434; max spread 422.875; stable=True
Task cancel (whole operation) / tiny: 297.965 [229.128-434.815] -> 275.905 [204.634-322.337] ms; gain 22.060; max spread 205.687; stable=False

Startup parent provider store reads / large: 73.037 [65.016-91.351] -> 30.918 [19.608-48.072] ms; gain 42.119; max spread 28.465; stable=True
Startup parent provider store reads / tiny: 16.168 [3.195-30.471] -> 31.902 [21.867-44.640] ms; gain -15.735; max spread 27.276; stable=False

Completion planning store reads / large: 1005.195 [860.989-1183.133] -> 15.673 [12.566-32.380] ms; gain 989.523; max spread 322.144; stable=True
Completion planning store reads / tiny: 35.329 [22.272-63.376] -> 16.288 [4.813-18.797] ms; gain 19.041; max spread 41.105; stable=False

Parent status-sync store reads / large: 64.630 [59.576-74.010] -> 4.729 [2.293-9.993] ms; gain 59.900; max spread 14.434; stable=True
Parent status-sync store reads / tiny: 9.730 [1.997-16.211] -> 2.812 [1.859-11.612] ms; gain 6.918; max spread 14.213; stable=False

Cancellation parent task store read / large: 16.606 [9.936-18.236] -> 0.719 [0.170-4.510] ms; gain 15.887; max spread 8.300; stable=True
Cancellation parent task store read / tiny: 0.764 [0.329-2.047] -> 0.215 [0.199-0.372] ms; gain 0.550; max spread 1.718; stable=False

Parent run interrupt-check store call / large: 11.719 [9.493-18.080] -> 0.646 [0.388-2.692] ms; gain 11.073; max spread 8.587; stable=True
Parent run interrupt-check store call / tiny: 1.102 [0.664-9.072] -> 0.788 [0.533-6.377] ms; gain 0.314; max spread 8.408; stable=False

base counts {"large": {"runs": 319, "nodes": 41392, "subagents": 754, "turn_items": 41712}, "tiny": {"runs": 19, "nodes": 92, "subagents": 54, "turn_items": 112}}
base behavior: {"measureDone": 18, "errors": [], "tools": {"delegate_task": 108, "task_status": 54, "task_cancel": 54}, "persistedTaskOutcomes": [{"prompt": "PERF:QUICK", "status": "completed", "result": "QUICK_OK", "n": 754}, {"prompt": "PERF:LONG", "status": "interrupted", "result": "LONG_READY", "n": 54}], "liveRunsAfterMeasurement": 0}
head counts {"large": {"runs": 320, "nodes": 41394, "subagents": 754, "turn_items": 41715}, "tiny": {"runs": 19, "nodes": 92, "subagents": 54, "turn_items": 112}}
head behavior: {"measureDone": 18, "errors": [], "tools": {"delegate_task": 108, "task_status": 54, "task_cancel": 54}, "persistedTaskOutcomes": [{"prompt": "PERF:QUICK", "status": "completed", "result": "QUICK_OK", "n": 754}, {"prompt": "PERF:LONG", "status": "interrupted", "result": "LONG_READY", "n": 54}], "liveRunsAfterMeasurement": 0}