There's no reason why the memory size limit should be expected to look like a magic number -- or even be the same number from one moment to the next. It will depend on the amount of free swap space, which might vary with how much file data is sitting around on /tmp.

update: I noticed that the NDBM module man page does talk about "errno 22", which happens when a given key and/or value being stored is too large (the limit is said to be 1008 bytes for key and value combined), though I guess that's not relevant here -- or maybe it is, hard to say based on the code in the OP.

Meanwhile, the "AnyDBM" man page cites "4k" as the "Size limit" for NDBM (hmm, I guess one of these two man pages might be off by about 3k), and it helpfully notes "?" for "Database Size"...