[BUG] ROUTING_TEST CVRPTW_Retail/4 SIGABRT — "Cost should improve" assertion in execute_vrp_moves

Describe the bug

ROUTING_TEST crashed with SIGABRT on the level0_retail/retail_float_test_t.CVRPTW_Retail/4 case. The cuopt_assert invariant on the aggregate cost change after a batch of VRP local-search moves did not hold:

ROUTING_TEST: cpp/src/routing/local_search/vrp/vrp_execute.cu:466:
bool cuopt::routing::detail::execute_vrp_moves(
    solution_t<i_t, f_t, REQUEST>&, move_candidates_t<i_t, f_t>&, i_t)
    [with i_t = int; f_t = float;
     cuopt::routing::request_t REQUEST = cuopt::routing::request_t::VRP]:
Assertion `cost_before - cost_after > EPSILON && "Cost should improve!"' failed.

./ci/run_ctests.sh: line 86: 4820 Aborted (core dumped) "${gt}" --gtest_output="xml:${xml_file}" "$@"
CRASH: ROUTING_TEST died from SIGABRT (exit code 134)

The assertion fires after a batch of accepted VRP moves is executed: the planner predicted a cost decrease, but the post-execution measured cost decrease was not greater than EPSILON. This indicates either stale move-candidate data, conflicting moves applied together, or a predicted/actual delta mismatch.

Where

• File: cpp/src/routing/local_search/vrp/vrp_execute.cu:466
• Function: cuopt::routing::detail::execute_vrp_moves<int, float, VRP>
• Test: level0_retail/retail_float_test_t.CVRPTW_Retail/4
• Binary: ROUTING_TEST (debug build — assertion is inside cuopt_func_call / cuopt_assert)

Reproducer / observed run

• PR: #1269 (CI-only change — does not touch routing code)
• Run: actions/runs/26237894930
• Job: conda-cpp-tests / 13.2.0, 3.13, amd64, ubuntu24.04, rtxpro6000, latest-driver, latest-deps (job 77225276444)
• Head SHA: 371edafa (branch pull-request/1269, base release/26.06)
• Surrounding tests CVRPTW_Retail/0..3 passed; only /4 aborted.

Expected behavior

execute_vrp_moves should only accept and execute moves whose net post-execution cost change matches the predicted delta within EPSILON. The assertion should hold for every batch on every accepted test instance.

Environment

• CI runner: rtxpro6000 (NVIDIA RTX PRO 6000 Blackwell Server Edition, 94.97 GiB VRAM)
• CPU: INTEL(R) XEON(R) GOLD 5512U, 16 threads
• CUDA: 13.2
• Driver: latest-driver
• Deps: latest-deps
• Build: conda, ubuntu24.04, amd64, Python 3.13
• Branch under test: release/26.06

Related issues

• #1221 (closed 2026-05-18) — similar CVRPTW_Retail SIGABRT, but on case /18 and root-caused to route-priority-sort out-of-bounds (fixed by 1d85aed6). Current failure is on /4 and at a different assertion site, so likely a distinct root cause that survived that fix.
• #866 (closed 2026-05-18) — same family of cost-invariant assertion but at local_search.cu:315 ("Cost mismatch after a move"), case /18. Cross-link for context.
• #402 (closed 2025-11-20) — older instance of the same "Cost should improve" assertion family.

Additional context

• PR #1269 modifies only .github/workflows/* and ci/release/update-version.sh, so the failure cannot have been caused by its changes. Either:
  1. The test is flaky on this hardware/CUDA combination, or
  2. A latent bug exists on release/26.06 that this run happened to expose.
• If flaky: consider whether cuopt_assert should be relaxed for the aggregated-batch case, or the failing instance saved for offline reproduction.
• If a real regression: bisect release/26.06 since #1221's fix landed (1d85aed6 on 2026-05-18) to identify the change that exposed it.