The current lru_cache code is complicated because it mixes several intentions. Let's untangle them.

The two goals are 1) in-memory LRU memoization to lessen remote data fetches; 2) persistent caching, primarily so that tests do not require network access.

Consequences of mixing these concerns are:

• Can't use external lru_cache code (incl. in 3.x)
• Configuration is confusing. uta connect() requires a cache mode that's different than the lru_cache mode, and neither checks whether the supplied value is legit.
• Learning mode is slow (probably because it writes back every time).
• As implemented, the hdp interface is also entangled in caching.

I'm going to park this as a placeholder for discussion. I think I'd like to see us revert to a cache-unaware UTA module, and a (new) caching hdp that accepts an underlying hdp (eg uta) that merely caches data.