CP-SAT solver takes way too long to shut down or does not stop at all despite given time limit

[S. K.]

I want to use v9.14, but it still has a pretty severe issue for me (problem 2 from this post).

The time limit I give is not respected.

I am using a CP-SAT solver with the following parameters:

random_seed: 2
use_optimization_hints: true
max_time_in_seconds: 540
log_search_progress: true
cp_model_probing_level: 1 (or 0, it does not matter)
max_presolve_iterations: 1
merge_no_overlap_work_limit: 1e9 (or 1e8, it does not matter)

hint_conflict_limit: 75000
log_subsolver_statistics: false
fix_variables_to_their_hinted_value: false

num_workers: 128
probing_deterministic_time_limit: 0.3

Here is a model for testing purposes:

model: test model

With 2 models similar to the testing model, I encountered the problem that the solver does not stop AT ALL. It runs until I get a memory overload and the whole process is killed.

(Output of on example where the solver does not stop at all: output)

(Similar but slightly less complex models do not have this problem, or not as bad. They take maybe 10-30 seconds longer to stop if max_time_in_seconds = 300 s.)

In general I have the feeling that the v9.14 needs a lot more memory than v9.10, which I am currently using.

I encountered the same problem but far less severe with v9.10.
For example: I start the solver with a solution callback that ends the solver after the first found solution. The first solution is found after 15.43 seconds, but the solver has a user/wall time of 182.827 seconds.
The parameters in this case were: 

Starting CP-SAT solver v9.10.4067
Parameters:
random_seed: 2
random_branches_ratio: 0.04
use_optimization_hints: true
max_time_in_seconds: 900
log_search_progress: true
search_branching: AUTOMATIC_SEARCH
cp_model_presolve: false
cp_model_probing_level: 1
max_presolve_iterations: 1
merge_no_overlap_work_limit: 1000000000
hint_conflict_limit: 75000
log_subsolver_statistics: false
fix_variables_to_their_hinted_value: true
num_workers: 12
probing_deterministic_time_limit: 0.3

Has anyone else experienced this? Are there any workarounds or parameter changes that might help with proper time limit enforcement?