Using SHAP on unseen data to understand model's predictions

[ETTAN93]

Assuming I have a model that is initialized as such:

model_estimator = LightGBMModel(
    lags=None,
    lags_past_covariates=[-3,-2,-1].
    lags_future_covariates=[-3,-2,-1],
    output_chunk_length=3
)

Question 1:
I then create the ShapExplainer object by fitting it to the training set.

shap_explain = ShapExplainer(model_estimator)
explanations = shap_explain.summary_plot()

Assuming now I want to use the ShapExplainer to explain data in the unseen test set, what should be defined as the foreground series?
I tried not providing the foreground series since target lags is 0 but it seems like that is not possible. In that case, should the foreground series, past_covariates and future covariates be the same series passed to the model.predict function, i.e. the foreground_series should end at the first prediction timestamp t=0?

shap_explain.explain(
    foreground_series = hf_data_dict['target_hf'][: test_set_start_date],
    foreground_past_covariates = hf_data_dict['past_cov_hf'][test_set_start_date: end_date],
    foreground_future_covariates = hf_data_dict['future_cov_hf'][test_set_start_date: end_date],
    horizons = [3]
)

Or should the foreground_series start from the test_set_start_date?

shap_explain.explain(
    foreground_series = hf_data_dict['target_hf'][test_set_start_date: ],
    foreground_past_covariates = hf_data_dict['past_cov_hf'][test_set_start_date: end_date],
    foreground_future_covariates = hf_data_dict['future_cov_hf'][test_set_start_date: end_date],
    horizons = [3]
)

Question 2:
I tried training the SHAP explainer on different background series (train or test set) but I am getting back the exact same SHAP results. For example

split_date = pd.to_datetime('2023-02-28 23:59:00')
test_set_start_date = pd.to_datetime('2023-03-01 00:00:00')

#shap explainer trained on train set but used to explain test set
shap_explainer = ShapExplainer(lgbm_model)

df1 = shap_explainer.explain(
    foreground_series = hf_data_dict['target_hf'][: split_date],
    foreground_past_covariates = hf_data_dict['past_cov_hf'][test_set_start_date: ],
    foreground_future_covariates = hf_data_dict['future_cov_hf'][test_set_start_date: ],
    horizons = [3]
)

#shap explainer trained on test set and used to explain test set
shap_explainer_test = ShapExplainer(
    lgbm_model,
    background_series = hf_data_dict['target_hf'][: split_date],
    background_past_covariates = hf_data_dict['past_cov_hf'][test_set_start_date: ],
    background_future_covariates = local_lgbm_hf_output.hf_data_dict['future_cov_hf'][test_set_start_date: ]
)

df2 = shap_explainer_test.explain(
    foreground_series = hf_data_dict['target_hf'][:split_date],
    foreground_past_covariates = hf_data_dict['past_cov_hf'][test_set_start_date: ],
    foreground_future_covariates = hf_data_dict['future_cov_hf'][test_set_start_date: ],
    horizons = [3]
)

#shap explainer trained on test set but changing foreground series to same as past and future cov
shap_explainer_test = ShapExplainer(
    lgbm_model,
    background_series = hf_data_dict['target_hf'][test_set_start_date: ],
    background_past_covariates = hf_data_dict['past_cov_hf'][test_set_start_date: ],
    background_future_covariates = local_lgbm_hf_output.hf_data_dict['future_cov_hf'][test_set_start_date: ]
)

df3 = shap_explainer_test.explain(
    foreground_series = hf_data_dict['target_hf'][:split_date],
    foreground_past_covariates = hf_data_dict['past_cov_hf'][test_set_start_date: ],
    foreground_future_covariates = hf_data_dict['future_cov_hf'][test_set_start_date: ],
    horizons = [3]
)

It seems like all 3 dataframes return exactly the same value which is very weird to me as the ShapExplainer is trained with a different background series. Also when looking at the base_values, they are also exactly the same. Is there a bug somewhere in the implementation or am I using the function wrongly?