That too is my experience, also with several K of effective burnt coins. It seems to go in cycles as far as I can make out. As I mentioned, the code is canonical (albeit probabilistic, AIUI).
There's possibly some further enlightenment to be had from the Technical Discussions doc that I added (basically just a collation of early bitcointalk posts but it is focused on minting/mining).

BTW, I have reworked ACME so that it now gets its data from SPARQL queries of the blockgraph and not from RPC commands sent to the client. It's is mostly working, just a few wrinkles to iron out. It's testnet-only atm, handling a few K blocks is way more tractable than 1.7million:

http://tessier.bel-epa.com:5064/test/

Work on browsing the blockgraph is continuing, as is the calculation of network hash stats, the ordered list of balances and the address/tx history.

Cheers,

Graham