This metric is bugged. For example, my big node lists "Expected time to share: 1.2 seconds", which is impossible to sustain, as the protocol-level minimum share difficulty will adjust upward until the network as a whole creates one share every 30 seconds.

A difficulty of 1520000 should result in an average of 1 share every 1520000 *2**48/0xffff = 6.53 PH. With a hashrate of 20 TH/s, that should take on average 5.44 hours.

On my to-do list is changing a few things with the difficulty-setting and share-orphaning algorithms to improve both fairness and also to assign large miners higher-difficulty shares. This should substantially improve the number of shares small miners mine. It will not increase expected revenue for small miners, though; it will only reduce the variance.

The averaging period that determines your revenue per block is approximately 3 days long, so you are expected to have around 14 shares in the share chain at steady state, and so your variance will be around 1/sqrt(14)~=27%.

Correct. On the other hand, if you find 2x as many shares as the expectation, then you get 2x the payout.

Yes, after about 3 days you will be back to ground zero. However, if p2pool finds a block 1.5 days after you stop mining, then you will still get paid at about 50% of your normal payout level. Your expected payout ramps up slowly over 3 days, and it ramps down slowly over 3 days.

Number of shares found on this node since it was (re)started.

That sounds weird. Perhaps you have your worker name/address set incorrectly? If that's the case, then you will end up mining to that node's default address, which usually belongs to the node operator. Your worker name needs to be a Bitcoin address starting with a 1 (no segwit or multisig addresses) and must not have an _ or ., though you may use + or / to set the pseudoshare or actual share difficulty.