If the difficulty can be directly observed by the script language, you
would need to re-evaluate all scripts in unconfirmed transactions
whenever the difficulty changes. This complicates implementation of
mempools, but it also makes reasoning about validity of (chains of)
unconfirmed transactions harder, as an unconfirmed predecessor may
have conditions that change over time.
To deal with potentially wildly varying difficulty, could the value exposed be the sum of accumulated PoW, or in other words the sum of each block's difficulty value in the entire chain? This should be a value that will only rise unless a reorg happens after a difficulty drop happens (only likely to be the result of users manually blacklisting an otherwise valid block that is several blocks back in the chain).
This mimics the effect of the block number which only grows. So if you're starting at time A with difficulty X, then you'd estimate what you think the accumulated PoW ought to be at time B with expected difficulty Y (as compared to the current value at time A), and put that value into the script.