Rik, thank you for going into this.  This surprised me indeed, too.
 

// compile with:
//
//  (clan)g++ -std=c++11 -Wall -Wextra -pedantic

#include <iostream>
#include <memory>
#include <vector>

std::unique_ptr<std::vector<int>> global_stuff;

std::vector<int> do_something () {
  std::vector<int> retval;

  return *global_stuff;

  return retval;
}

int main () {
  global_stuff = std::unique_ptr<std::vector<int>> (new std::vector<int> ({1, 2, 3}));

  std::vector<int> val = do_something ();

  for (auto const& v : val) {
    std::cout << v << ' ' << std::endl;
  }

  return 0;
}

Entirely undetected my garbage in both GCC and clang?!

Anyways, I fixed this
https://hg.savannah.gnu.org/hgweb/octave/rev/bc521cd24922

It might take more time to get into the rest, where to fix this appropriately.

Taking a look at the code in ov-struct.cc for cell2struct I see

  if (! (args(1).iscellstr () || args(1).is_char_matrix ()))
    error ("cell2struct: FIELDS must be a cell array of strings or a character matrix");

  int dim = 0;

  if (nargin == 3)
    {
      if (! args(2).is_real_scalar ())
        error ("cell2struct: DIM must be a real scalar");

      dim = args(2).int_value () - 1;
    }

  if (dim < 0)
    error ("cell2struct: DIM must be a valid dimension");

  const Cell vals = args(0).cell_value ();
  const Array<std::string> fields = args(1).cellstr_value ();

  octave_idx_type ext = (vals.ndims () > dim ? vals.dims ()(dim) : 1);

  if (ext != fields.numel ())
    error ("cell2struct: number of FIELDS does not match dimension");

So, the problem is a little deeper.  Have to look at cellstr_value() and see how it operates.

A subtle truncation happens when calling

containers.Map(['foo';'bar';'baz'], [1,2])

https://lists.gnu.org/archive/html/octave-maintainers/2020-06/msg00088.html

The cause is that cell2struct allows the following:

Octave 6.0.1:

>> cell2struct({[1,2]}, {['foo','bar','baz']})
ans =

  scalar structure containing the fields:

    foobarbaz =

       1   2


>> cell2struct({[1,2]}, {['foo';'bar';'baz']})
ans =

  scalar structure containing the fields:

    foo =

       1   2

while Matlab R2020b does not:

>> cell2struct({[1,2]}, {['foo','bar','baz']})

ans =

  struct with fields:

    foobarbaz: [1 2]

>> cell2struct({[1,2]}, {['foo';'bar';'baz']})
Error using cell2struct
Field name must be a character vector or string scalar.

I agree to Matlab's behavior in this case, because my valuable key is truncated, without a warning.