Augment & Forecast occasionally does not produce consistent mean for TSLM(log(mr) ~ trend()) model

In the below example, using a training set of 2010:2019 produces the following chart:

Note the break in the forecasted mean.

This does not happen when either setting training to 2015:2019 or just using lin. trend model: TSLM(mr ~ trend())

Example:

library(fable)
library(tsibble)
library(dplyr)

df <- tibble(
  date = 2010:2023,
  mr = c(704, 852, 935, 520, 750, 305, 560, 769, 774, 703, 941, 439, 912, 584)
)

df_train <- df |> filter(date %in% 2010:2019)
df_test <- df_bl |> filter(year > 2019)

mdl <- df_train |>
  as_tsibble(index = date) |>
  model(lm = TSLM(log(mr) ~ trend()))

bl <- mdl |>
  augment() |>
  rename(.mean = .fitted)
fc <- mdl |> forecast(h = 4)

df_plot <- bind_rows(bl, fc) |> select(date, .mean)

df_plot |> autoplot(.vars = .mean)

Start by running the supplied R example, checking the augment() and forecast(h = 4) outputs for TSLM(log(mr) ~ trend()) trained on 2010:2019. Trace the code paths producing .fitted and .mean, then verify that combining the fitted values and forecasts produces no break in the forecasted mean while the linear-trend case remains consistent.