I understand that one can translate a nucleotide sequence and run PSI-BLAST on the protein (proteins if you take the 6 reading frames), but I'm looking for distant homology for bacterial small RNAs (typically 50-200 nucleotides long and noncoding).

If there is no such resource, what are the main obstacles to this implementation?

PSI-BLAST creates a protein profile of similar proteins (essentially a scoring matrix based on multiple sequence alignments) that are then used for further DB searches. On the nucleotide level this simply does not make much sense, especially for short sequences.

For one, there is much less potential variability which is probably worse for non-coding sequences (although for certain sRNAs one could expect the presence of specific anti-codons). Any profile generated from there is likely to yield a massive amount of random hits.