Kafka to Paimon repeated measurements

Query	Variant	Stock median [range], s	Native median [range], s	Stock/native
q0	exact	1.893 [1.866–1.937]	1.270 [1.248–1.471]	1.49×
q1	approximate decimal (incompatible)	1.826 [1.688–1.835]	1.222 [1.179–1.533]	1.49×
q1	exact	1.928 [1.893–2.232]	1.326 [1.191–1.333]	1.45×
q2	exact	1.394 [1.388–1.417]	1.158 [1.131–1.159]	1.20×
q3	exact	1.665 [1.389–2.321]	1.504 [1.302–1.794]	1.11×
q4	exact	11.905 [6.190–24.370]	5.376 [5.138–13.164]	2.21×
q5	exact	3.583 [3.334–3.972]	4.618 [4.555–4.674]	0.78×
q7	exact	6.382 [5.653–25.175]	3.963 [3.908–11.540]	1.61×
q8	exact	2.774 [1.557–3.787]	2.097 [2.012–2.265]	1.32×
q9	exact	27.723 [7.297–27.816]	6.954 [6.744–18.020]	3.99×
q10	exact	1.948 [1.858–1.950]	1.435 [1.425–1.454]	1.36×
q10	native datetime (incompatible)	1.859 [1.826–1.970]	1.433 [1.423–1.438]	1.30×
q11	exact	5.345 [5.270–5.473]	1.057 [1.055–1.070]	5.06×
q12	exact	1.578 [1.540–1.631]	1.182 [1.177–1.182]	1.33×
q13	exact	1.600 [1.582–1.630]	1.222 [1.160–1.238]	1.31×
q14	exact	1.743 [1.705–2.167]	1.476 [1.445–1.652]	1.18×
q14	native datetime (incompatible)	1.721 [1.705–2.181]	1.431 [1.430–1.439]	1.20×
q15	exact	9.680 [9.506–9.709]	5.839 [5.768–5.875]	1.66×
q15	native datetime (incompatible)	9.567 [9.530–9.984]	6.013 [5.815–6.317]	1.59×
q16	exact	6.930 [6.817–10.617]	5.090 [4.784–8.721]	1.36×
q16	native datetime (incompatible)	7.090 [6.822–7.152]	4.960 [4.897–4.968]	1.43×
q17	exact	4.003 [3.748–4.300]	2.112 [2.078–2.706]	1.90×
q17	native datetime (incompatible)	5.510 [4.390–7.756]	3.925 [1.469–4.957]	1.40×
q18	exact	18.688 [5.767–24.520]	11.560 [3.810–20.835]	1.62×
q19	exact	17.964 [5.928–28.321]	9.392 [3.582–18.927]	1.91×
q20	exact	4.359 [4.137–14.932]	9.493 [4.236–20.648]	0.46×
q21	exact	1.595 [1.549–1.618]	1.453 [1.445–1.468]	1.10×
q21	native regex/case (incompatible)	1.608 [1.559–2.054]	1.099 [1.088–1.114]	1.46×
q22	exact	1.765 [1.705–2.333]	1.182 [1.172–1.183]	1.49×
q23	exact	10.881 [10.321–22.122]	17.014 [11.059–19.158]	0.64×

Exact SQL geomean: 1.43×; README selection: 1.43×.

One ratio per query; medians use every measured repetition. All warmups and slow samples remain in observations.jsonl.