When the current locale doesn't match the current voice, grep would be
surprised by the produced output and believe that this is not text, for
instance with LC_ALL=ru_RU.CP1251 we get:

  TEST      tests/language-replace.test
[...]
testing mk
grep: (standard input): binary file matches
2d1
< Translate 'пејзаж'

But we can give -a to grep so it always considers its input as text.

Now that all errors are fixed.

Otherwise we would miss errors produced on shutdown.