nginx will so log them in standard error file, mainly to complain headers are too big. Client can then get a 502 error, as demonstrated here on Jenkins.
Plan is to configure php-fpm more like the webserver-legacy role: log errors in a log file instead of sending stacktrace to the client. For some applications, depending of what they output when there is an error, it could also pose a (dubious as errors output should be sanitized but real) security issue.
Spin-off from T1731.