Negative % variance explained for Cross-validated Hierarchical Generalized Additive Model

Question

So I have a dataset I've split into training & test with an 80/20 ratio. It's a hieararchically structured dataset, individual human participants each with many observations, so when splitting I ensure that there's an 80/20 split within each participant. I fit to the training data via (excuse my idiosyncratic code-style):

fit = mgcv::bam(
    formula = power ~ 
        (
            te(
                lat , long , cov
                , d = c(2,1)
                , bs = c("sos","cr")
                , k = c(32,10) #max supported by num unique
            ) 
            + te(
                lat , long , cov, participant
                , d = c(2,1,1)
                , bs = c("sos","cr","re")
                , k = c(32,10,40) #max supported by num unique
            )
        )
    , data = training_data
    , method = "fREML"
    , discrete = TRUE
    , chunk.size = 0
    , nthreads = parallel::detectCores()
    , gc.level = 0
)

Which seems to converge just fine. Then I get the predictions for both the training (as a sanity-check) and testing datasets, once with predictions sensitive to the term that includes Participant as a random effect, and once with predictions sensitive only to the group-level terms, then compute the prediction error for each and finally compute the ratio of said prediction errors to the sum original variance (computed by hand as a sanity check):

(
    both_data #has both training and testing datasets, indicated by column `dataset`
    %>% dplyr::mutate(
        re_term_present = 
            (
                mgcv::predict.bam(
                    fit
                    , newdata = both_data
                    , discrete = T
                    , n.threads = parallel::detectCores()
                )
            )
        , re_term_absent = 
            (
                mgcv::predict.bam(
                    fit
                    , newdata = both_data
                    , discrete = T
                    , n.threads = parallel::detectCores()
                    , exclude = 'te(lat,long,cov,participant)'
                )
            )
    )
    %>% tidyr::pivot_longer(
        cols = c(re_term_present,re_term_absent)
        , names_to = 're_term'
        , values_to = 'model_prediction'
    )
    %>% dplyr::group_by(
        dataset
        , re_term
    )
    %>% dplyr::mutate(
        model_squared_error = (power-model_prediction)^2
        , null_squared_error = (power-mean(power))^2
    )
    %>% dplyr::summarise(
        model_sse = sum(model_squared_error)
        , null_sse = sum(null_squared_error)
        , model_null_ratio = model_sse/null_sse
        , prop_accounted = 1-model_null_ratio
    )
)

Yielding the table:

  dataset re_term         model_sse null_sse model_null_ratio prop_accounted
  <chr>   <chr>               <dbl>    <dbl>            <dbl>          <dbl>
1 test    re_term_absent    139259.  105291.            1.32          -0.323
2 test    re_term_present    86585.  105291.            0.822          0.178
3 train   re_term_absent    590409.  454435.            1.30          -0.299
4 train   re_term_present   300710.  454435.            0.662          0.338

The result for predictions that take into account the random effects of participants (re_term_present) make sense: predictions for both test and training account for a positive proportion of the variance of the observed data, with test predictions achieving a lower proportion-accounted-for than the training predictions.

But I can't wrap my head around the result for predictions that don't take into account the random effects of participants (re_term_absent); how can the model predictions be adding variance? I thought the term that includes Participants as a random effect would be modelled as deviations centered on the mean-across-participants term, in which case I'd expect maybe as low as 0% variance accounted-for if said mean term were indeed zero, but certainly not negative values. How am I thinking about this incorrectly?