[BUG] Routing solver: break-dimension bugs

Describe the bug

For the entire problem setup and reproduce, see repro_break_dim_bugs.py.

Both stem from a single design assumption: every route has exactly N breaks. The cost formula max(0, breaks − N) only penalizes excess; squeeze unconditionally inserts every dim; eject_until_feasible protects breaks; find_break_insertions only moves existing breaks.

Observed

expect status 1 since the problem is infeasible
  A: cost=20.00 vehicles=1
      vehicle 0: 0(Depot) -> 1(Delivery) -> 0(Depot)
expect status 0 and no break in solution since it is not needed
  B: cost=410.25 vehicles=1
      vehicle 0: 0(Depot) -> 2(Break) -> 1(Delivery) -> 0(Depot)

Expected

expect status 1 since the problem is infeasible
  A: INFEASIBLE (status=1)
expect status 0 and no break in solution since it is not needed
  B: cost=20.00 vehicles=1
      vehicle 0: 0(Depot) -> 1(Delivery) -> 0(Depot)

Why it matters

This issue negatively affects UX and is a blocker for feature development.
Case A is a bug with straightforward fix.
Case B appears to be a solver/design limitation rather than a simple feasibility bug.

Environment

• cuOpt: b311a990 (main), built from source (Release).
• Single A100 80 GB PCIe
• Driver Version: 580.65.06, CUDA Version: 13.0, nvcc V13.1.115
• Linux 5.15.179, Python 3.14, miniconda.

Notes

• Adding distance breaks semantics forces EVs travel around chargers to comply with the required breaks due to issue B.
• To fix case A override inf_cost[BREAK] in route_t::compute_cost with max(0, required - breaks) using a dynamic required-count derived from the route's actual arrival_at_depot. Squeeze still inserts every dim so no construction path changes; all 83 routing tests pass.
• an attempt to add in-LS eject-only candidate to fix case B perturbed LS convergence (heterogenous_breaks became flaky).