NAs causing failures in training combination models

I have a combination model that's running into some problems with NAs.

The data looks like this (see end of message for actual dput format to recreate the exact data set):

Browse[2]> summary(t1)
 interval_time                     dalmp           target       
 Min.   :2023-08-29 23:00:00   Min.   :21.89   Min.   :-1.8695  
 1st Qu.:2023-08-30 06:15:00   1st Qu.:24.90   1st Qu.:-0.1650  
 Median :2023-08-30 13:30:00   Median :34.82   Median : 0.1793  
 Mean   :2023-08-30 13:30:00   Mean   :40.45   Mean   : 0.2144  
 3rd Qu.:2023-08-30 20:45:00   3rd Qu.:53.67   3rd Qu.: 0.7105  
 Max.   :2023-08-31 04:00:00   Max.   :81.69   Max.   : 1.5769  
                                               NA's   :2        
Browse[2]> t1
# A tsibble: 30 x 3 [1h] <UTC>
   interval_time       dalmp target
   <dttm>              <dbl>  <dbl>
 1 2023-08-29 23:00:00  64.8  1.58 
 2 2023-08-30 00:00:00  55.0  1.57 
 3 2023-08-30 01:00:00  49.6 NA    
 4 2023-08-30 02:00:00  37.0 NA    
 5 2023-08-30 03:00:00  34.1  0.217
 6 2023-08-30 04:00:00  27.9  0.111
 7 2023-08-30 05:00:00  24.8 -0.132
 8 2023-08-30 06:00:00  24.6 -0.121
 9 2023-08-30 07:00:00  22.8 -0.195
10 2023-08-30 08:00:00  22.1 -0.328
# ℹ 20 more rows
# ℹ Use `print(n = ...)` to see more rows

And the training goes like this:

rt_fit <- fabletools::model(training,
                              cmbn1 = combination_model(
                                ARIMA(target ~  dalmp),
                                NNETAR(target ~  dalmp),
                                AR(target ~ order(2)),
                                cmbn_args = list(weights = "inv_var")),
                              .safely = FALSE)

# Error in qr.default(XX) : NA/NaN/Inf in foreign function call (arg 1)

I thought I might be able to work around this by deleting the rows with NAs, but it doesn't seem happy with that either:

Error in check_gaps(.data) : 
  .data contains implicit gaps in time. You should check your data and convert implicit gaps into explicit missing values using `tsibble::fill_gaps()` if required.

Is there a problem in combination_model that's causing this? Any suggested workaround? Any help greatly appreciated!

I'm using:

> R.version
               _                           
platform       aarch64-apple-darwin20      
arch           aarch64                     
os             darwin20                    
system         aarch64, darwin20           
status                                     
major          4                           
minor          2.2                         
year           2022                        
month          10                          
day            31                          
svn rev        83211                       
language       R                           
version.string R version 4.2.2 (2022-10-31)
nickname       Innocent and Trusting      

> packageVersion('fabletools')
[1] ‘0.3.3’
> packageVersion('fable')
[1] ‘0.3.3’

Here's the input data for reproduction purposes:

training <- 
structure(list(interval_time = structure(c(1693350000, 1693353600, 
1693357200, 1693360800, 1693364400, 1693368000, 1693371600, 1693375200, 
1693378800, 1693382400, 1693386000, 1693389600, 1693393200, 1693396800, 
1693400400, 1693404000, 1693407600, 1693411200, 1693414800, 1693418400, 
1693422000, 1693425600, 1693429200, 1693432800, 1693436400, 1693440000, 
1693443600, 1693447200, 1693450800, 1693454400), tzone = "UTC", class = c("POSIXct", 
"POSIXt")), dalmp = c(64.772, 55.031, 49.5954, 37.0346, 34.122, 
27.9416, 24.8358, 24.6128, 22.81, 22.1077, 21.8885, 21.9106, 
22.9985, 25.0922, 25.7776, 30.3662, 32.1598, 35.5114, 39.2446, 
44.2576, 59.0444, 65.3967, 76.7988, 81.6873, 73.1935, 64.3783, 
45.2286, 36.5493, 26.4268, 22.6344), target = c(1.57691363525177, 
1.56855930002442, NA, NA, 0.216786192700598, 0.110616995970005, 
-0.131849633308032, -0.12072764832231, -0.195075151428118, -0.327716438359099, 
-0.403658249907274, -0.361481112249875, -0.254285754950902, -0.154945697500086, 
-0.142376837846178, 0.141843887915466, 0.376564223792779, 0.327934476018599, 
0.875235233052582, 0.550416417555717, 0.682786557929835, 0.482133615075935, 
0.952303345137791, 1.16381971509555, 0.793706899407923, 0.854539241873788, 
0.239481464973916, -0.0814464383521666, -1.86950100289322, -0.867599159975814
)), class = c("tbl_ts", "tbl_df", "tbl", "data.frame"), row.names = c(NA, 
-30L), key = structure(list(.rows = structure(list(1:30), ptype = integer(0), class = c("vctrs_list_of", 
"vctrs_vctr", "list"))), class = c("tbl_df", "tbl", "data.frame"
), row.names = c(NA, -1L)), index = structure("interval_time", ordered = TRUE), index2 = "interval_time", interval = structure(list(
    year = 0, quarter = 0, month = 0, week = 0, day = 0, hour = 1, 
    minute = 0, second = 0, millisecond = 0, microsecond = 0, 
    nanosecond = 0, unit = 0), .regular = TRUE, class = c("interval", 
"vctrs_rcrd", "vctrs_vctr")))

Reproduce the failure using the supplied training object and the model() call with combination_model(), ARIMA(), NNETAR(), and AR(). Start by tracing how missing target values and gaps are handled, including the fill_gaps() path, before the qr.default() error. Done means the behavior with these data and documented missing-value handling is understood and a reliable workaround or fix is identified.