Hi book_woorm,
I have never been able to find a definitive answer but I have noticed that the length of time before lockup/crashes is maximised by not using DNS lookups so I use hard coded IP addresses.
Its not good as a long term fix but it makes a crash rate of hours turn into days.
The only section I still use a lookup is once every 24hrs I do a lookup for a timeserver at oceania.pool.ntp.org. If my home router had a timeserver function I could even hard-code that IP addr as well.
I have read many theories about the reasons for this, lack of memory often seems to come up as a possibilty. (have you seen the Goldilocks board?, 16K ram!)
To get around my lockups, I use the watchdog timer to reboot the board. It always seems to come back up OK.
On my Arduino server, I have an uptime timer (days/hrs/min/sec) so when I look at the page I know how long its been since its last reset.
What I dont differentiate (yet) is whether a reboot is due to an ethernet lockup or a string of five consecutive logging-service post failures. I will remedy that soon.
It seems that the logging sites can miss quite a few posts in busy times, at least thingspeak can, nimbits seems more reliable in that sense.
This is only my personal experiences.
One thing you are doing that i dont is writing to the SD card. I just read from one. I'm not sure if writing uses more ram.
Good luck with your coding, I hope you find some answers 
Stewie