Profile CI and shrink turnaround time

[ramakrishnap-nv]

Discovery phase first (repeated downloads, test-tree pruning on change, etc.), then scope + implement incrementally.

Already done (post-26.06 baseline):

• Removed dataset downloads from Python tests (ci: remove dataset downloads from Python test scripts #1458)
• pytest-xdist + parallel test runs (ci(test): enable pytest-xdist for cuopt Python tests #1486, ci(test): fix grpc server GPU-memory leak and run cuopt tests in parallel #1512)
• Reduced wheel-tests-cuopt-server matrix (ci: reduce wheel-tests-cuopt-server matrix to 1 job per CUDA major #1457)
• Removed lean CI matrix (ci: remove lean CI matrix from pr.yaml #1474)

Remaining: measure current turnaround, find remaining repeated downloads, complete test-tree pruning on change.

Owner: — · Size: XS (discovery)

[ramakrishnap-nv]

Related open PR:

• tests: retire correctness/regression tests covered by C++ gtests #1461 — tests: retire correctness/regression tests covered by C++ gtests (reduces test runtime / turnaround)

[ramakrishnap-nv]

Related closed (unmerged) experiments — xdist worker-count tuning not yet landed:

• ci(test): experiment with -n 8 xdist workers #1500 — experiment with -n 8 xdist workers
• ci(test): increase xdist workers from 2 to 8 #1501 — increase xdist workers from 2 to 8