Hi Brandon - good questions, see below

Is there any accepted standard for what constitutes a "high" versus a "low" value for any of the following?

1) singular values

Singular values in PLS are essentially covariances so are scale-dependent (cf. correlation) so 'high' vs 'low' requires that comparisons to involve the same data.

2) probability of permuted SV being higher than observed

here you can use a 'traditional' alpha level to assess the likelihood of the observed singular values given random permutations of the data (e.g., P<0.05 indicates a random sample produces a singular great than or equal to the observed only 5% of the time - i.e., 5% probability the original obs is chance)

3) percent cross-block covariation

like #1, this depends on the data set.  You can compare %'s within experiments but not between unless the designs are comparable.  For clarity, consider a simple design with two conditions.  A "task" PLS on these data would produce 1 latent variable accounting for 100% of the cross-block covariation.