Drop a misbehaving paired baseline instead of failing the run

[martian56]

Follow up to #69. The paired baseline server failed its setup statement on the arm64 profile and failed the whole bench job, which meant #71 could not be measured at all.

The paired baseline is an optimisation on measurement quality, not a requirement. If its server or its loadgen misbehaves the run now discards it and the report falls back to the cross run baseline, which is what happened before #69 anyway.

[github-actions]

Bench report

✅ Best improvement: linux-x64-full mixed at +19.6%. Nothing regressed beyond noise.

Baseline: main @ same runn. Cells read baseline → this PR. Verdicts need more than ±15% to leave the noise band.

linux-arm64-full

workload	ops/s	Δ ops/s	read p99 ms	write p99 ms	write p99.9 ms
reads	30378 → 30577	⚪ +0.7%	0.23 → 0.23	-	-
mixed	13698 → 12641	⚪ -7.7%	0.51 → 0.43	12.46 → 20.83	30.85 → 31.82
writes	5495 → 5963	⚪ +8.5%	-	9.32 → 7.48	29.17 → 26.19

linux-x64-1cpu-1gb

workload	ops/s	Δ ops/s	read p99 ms	write p99 ms	write p99.9 ms
reads	23930	new	0.22	-	-
mixed	6221	new	0.34	39.43	278.54
writes	2051	new	-	31.46	182.58

linux-x64-2cpu-2gb

workload	ops/s	Δ ops/s	read p99 ms	write p99 ms	write p99.9 ms
reads	17424	new	0.38	-	-
mixed	11901	new	0.64	3.23	26.00
writes	6190	new	-	1.94	22.64

linux-x64-full

workload	ops/s	Δ ops/s	read p99 ms	write p99 ms	write p99.9 ms
reads	63521 → 63949	⚪ +0.7%	0.11 → 0.11	-	-
mixed	15698 → 18767	🟢 +19.6%	0.14 → 0.13	24.48 → 19.21	58.63 → 40.18
writes	6212 → 5950	⚪ -4.2%	-	2.46 → 9.70	60.95 → 38.69

windows-x64-full

workload	ops/s	Δ ops/s	read p99 ms	write p99 ms	write p99.9 ms
reads	22404 → 22344	⚪ -0.3%	0.35 → 0.35	-	-
mixed	6002 → 5890	⚪ -1.9%	0.34 → 0.35	23.73 → 24.97	38.33 → 37.20
writes	772 → 782	⚪ +1.3%	-	16.21 → 16.42	25.20 → 25.77

Field reference

All engines in identical containers, same driver, measured 2026-07-24 @ ba0ec39. A different harness than the tables above, so read this as standing, not as this PR's delta.

linux-arm64-full

engine	reads	mixed	writes
quantadb	17978	12420 (27ms wp99)	6449 (19ms wp99)
postgres18	15227	14723 (2ms wp99)	9923 (1ms wp99)
mariadb114	4221	4437 (8ms wp99)	6666 (3ms wp99)
mysql84	4258	4394 (9ms wp99)	4274 (4ms wp99)

linux-x64-1cpu-1gb

engine	reads	mixed	writes
quantadb	19419	10854 (29ms wp99)	5170 (39ms wp99)
postgres18	10196	8580 (43ms wp99)	5912 (46ms wp99)
mariadb114	3455	3732 (4ms wp99)	5626 (8ms wp99)
mysql84	3497	3628 (4ms wp99)	3259 (36ms wp99)

linux-x64-2cpu-2gb

engine	reads	mixed	writes
quantadb	22528	16774 (5ms wp99)	8374 (4ms wp99)
postgres18	17060	17006 (2ms wp99)	11404 (1ms wp99)
mariadb114	3735	4006 (3ms wp99)	6741 (3ms wp99)
mysql84	3804	3845 (4ms wp99)	4847 (3ms wp99)

linux-x64-full

engine	reads	mixed	writes
quantadb	18144	15075 (6ms wp99)	8025 (5ms wp99)
postgres18	15402	15154 (2ms wp99)	10872 (1ms wp99)
mariadb114	3437	3699 (4ms wp99)	6138 (3ms wp99)
mysql84	3490	3615 (4ms wp99)	3996 (5ms wp99)

Shared CI runners: treat single-profile swings as suggestive, agreement across profiles as real.