[lmi] An unreasonably effective defect mitigation

[Greg Chicares]

There's a defect on branch valyuta/004 right now (SHA1 3742c5405d9):
    ,scale_fwd_         {detail::perform_pow(max_prec_real(10.0), decimals)}
    ,scale_fwd_c_       {detail::perform_pow(max_prec_real(10.0), decimals - 
currency::cents_digits)}
The first quoted line has been in 'master' for decades. The second
quoted line is new on this branch, and it's defective: what if
'decimals' is so negative that subtracting two from it (which is what
  decimals - currency::cents_digits
does) causes (signed) integer underflow? I understand that, and can
take whatever steps seem necessary to prevent anyone from rounding to
INT_MIN-2 digits. (As a practical matter, I'll probably just revise
the unit test that triggered this.)

But here's what I found surprising. Along my journey of discovery that
led to that insight, I used my favorite debugging technique (adding
statements to print internal state on stdout) and found that doing so
made the problem vanish--even when I reduced the added statements to
  std::cout << std::flush;
. My analysis is that subtracting two from INT_MIN is UB, one possible
consequence of which is that a failing unit test stops failing. Another
is that demons don't fly out your nose when it seems they should. This
just feels so remarkable that I was moved to share it.

To reproduce:

/opt/lmi/src/lmi[0]$LMI_COMPILER=gcc ; LMI_TRIPLET=x86_64-pc-linux-gnu ; . 
/opt/lmi/src/lmi/set_toolchain.sh         
/opt/lmi/src/lmi[0]$git switch valyuta/004
Already on 'valyuta/004'
Your branch is up to date with 'origin/valyuta/004'.
running post-checkout hook...
...post-checkout hook finished.
/opt/lmi/src/lmi[0]$git status
On branch valyuta/004
Your branch is up to date with 'origin/valyuta/004'.

nothing to commit, working tree clean
/opt/lmi/src/lmi[0]$git rev-parse HEAD
3742c5405d96393163dc45e8ed9f743d673c2def

/opt/lmi/src/lmi[0]$make $coefficiency unit_tests 
unit_test_targets=round_to_test
[...]
???? test failed: Caught exception of type
    'St11logic_error'
  when type
    'St12domain_error'
  was expected.
[file /opt/lmi/src/lmi/round_to_test.cpp, line 565]
[...]
???? 1 test errors detected; 8486 tests succeeded

Now the mystery: that error vanishes if we add a trivial
trace statement for debugging...and, in this minimal case,
even if that statement does pretty much nothing:

sed -i round_to.hpp -e'/<limits>/s/^/#include <iostream>\n/' -e '/return 
max/s/^/        std::cout << std::flush;\n/'
/opt/lmi/src/lmi[0]$make $coefficiency unit_tests 
unit_test_targets=round_to_test               
[...]
.... 8487 tests succeeded
!!!! no errors detected