transnoentities as this method send back UTF-8. Result was that CSV file
was generated empty (tested on Unix Ubuntu workstation).
Replaced now with $conf->file->character_set_client so it works for me
(using UTF-8 by default.
To complete the fix my feeling is that we should not force charset till
we are not writing the final result. But needs to be confirmed by a
senior Dolibarr developer.