[BUG] Flaky: DeterministicBB deterministic_across_runs/swath1_threads8 — run 3 diverges under work_limit #1793

DeterministicBB/DeterministicBBInstanceTest.deterministic_across_runs/swath1_threads8
fails intermittently in conda-cpp-tests. 1 of 528 tests; the rest pass.

Seen on run 32773153161
(13.3.0, 3.13, arm64, ubuntu26.04, l4, latest-driver, latest-deps). Surfaced on
#1763, which changes only .github/workflows/*.yaml for wheel publishing, so it
is not caused by that PR.

What happened

Params: ("/mip/swath1.mps", 8, 60, 4) — 8 CPU threads, time_limit=60,
work_limit=4. Seed for this occurrence: 480747044 (the test draws from
std::random_device per run and prints it, so each CI run exercises a different
seed).

The test solves the same problem three times with the same seed and compares.
Runs 1 and 2 agreed exactly; run 3 diverged:

Zero Run 1 vs 2 assertion failures, 128 Run 1 vs 3 failures.

determinism_test.cu:217: Failure
  solution1.get_objective_value()  Which is: 382.30780375000012
  solution3.get_objective_value()  Which is: 384.26733044999986

determinism_test.cu:220: Failure
  solution1.get_solution_bound()   Which is: 341.0518313
  solution3.get_solution_bound()   Which is: 341.16572324999987

Why this looks like a real bug rather than tolerance

All three runs terminated on the work limit, not the clock:

Work limit reached. Stopping the solver...
Explored 986 nodes (10364 simplex iterations) in 3.55s.
Best objective 3.842673e+02, best bound 3.411657e+02, gap 11.22%.

work_limit exists precisely so termination does not depend on wall time. Three
runs with identical settings and seed consuming different amounts of work before
the counter trips means the divergence is upstream of termination — the search
itself took a different path — rather than a formatting or floating-point
tolerance question. The differing node and simplex-iteration counts point the
same way.

The failure is also not marginal: variable 1 differs by 2.0
(1.78e-14 vs 2.0) and variable 11 by 1.0 (1.0 vs 0), i.e. different
integer assignments, not last-bit noise.

Notes

• Reproducing needs the seed pinned; the test's use of std::random_device
  means a rerun exercises a different seed and will often pass.
• Two of three runs matching suggests a race that usually resolves the same way
  — consistent with 8 threads and CUOPT_MODE_DETERMINISTIC.
• [BUG] Flaky Determinstic B&B test #1134 ([BUG] Flaky Determinstic B&B test, closed 2026-04-24) was the same
  test with a different instance (seymour1_threads16), so this class of
  failure has recurred.
• Related open work: [FEA] Determinism for MILP heuristics #144, [FEA] MIP: Implement Full Determinism (Heuristics Sync) #882, [FEA] End-to-End Determinism #1058, [FEA] Infra: Implement Infrastructure/Checks for Determinism #883, [FEA] Heuristics should use work estimates instead of time limits #1069.