Fix flaky reasoning budget tests on Metal by using greedy sampling #419

Job	Run time
Start gate (abort window)	3s
Code style (spotless) + package graph	47s
C++ Tests Ubuntu Latest x86_64	5m 47s
Build WebUI assets (shared)	1m 17s
Test (vmlens interleavings)	1m 8s
Cross-Compile manylinux_2_28 x86_64 (CUDA)	14m 51s
Build and Test Windows 2025 x86_64 (VS 2026)	16m 19s
Cross-Compile Android aarch64	2m 6s
Cross-Compile manylinux2014 x86_64	1m 50s
Build and Test macOS 14 arm64 (Metal)	2m 25s
Build and Test macOS 15 arm64 (no Metal)	2m 35s
Build and Test Linux aarch64	2m 58s
Cross-Compile Android aarch64 (OpenCL/Adreno)	1m 59s
Build and Test macOS 15 arm64 (Metal)	2m 29s
Build and Test Windows 2025 x86 (Ninja Multi-Config + sccache, eval)	3m 51s
Build and Test Windows 2025 x86_64 (Ninja Multi-Config + sccache, eval)	4m 8s
Build and Test Windows 2025 x86 (VS 2026)	16m 35s
Java Tests Ubuntu Latest x86_64	13m 41s
Java Tests macOS 14 arm64 (Metal)	18m 44s
Java Tests macOS 15 arm64 (Metal)	22m 59s
Java Tests macOS 15 arm64 (no Metal)	19m 58s
Java Tests Windows 2025 x86_64 (Ninja, eval)	17m 23s
Java Tests Windows 2025 x86_64 (VS 2026)	15m 42s
Package JARs	1m 46s
Report	25s
Check: v* tag	-1s
Check: main branch / SNAPSHOT	-1s
Publish Snapshot to Central	-1s
Publish Release to Central	-1s
Update Snapshot Pre-release on GitHub	0s
Attach Signed Binaries to GitHub Release	0s
	3h 11m 42s

Provide feedback