Apparently the Pentaho logging behaviour is a compromise between logging completeness and application performance. In addition, I think the status applet queries the log4j buffer for it's current state. TO get the most complete logs, set up a logging file on the job/trans executers at the top level (in the step/entry logging tab as in the screen shot)
Of course this means you can't read it via kettle/status. You could set up a pentaho transformation to read the file and return it as text output.