I was able to replicate what you've done to import the microgreens data and build a 23S classifier. So, I think you did everything correctly.
Not necessarily. DADA2, and other tools, perform de novo chimera detection. You can use any marker gene you'd like. To better optimize chimera detection and removal checkout this post.
The only thing I can think of, is that your reads might be in a mixed or reverse orientation. That is qiime feature-classifier fit-classifier-naive-bayes requires your data be in the same orientation as the classifier. You could run:
qiime rescript orient-seqs \
--i-sequences your-denoised-sequence-data.qza \
--o-oriented-seqs oriented-seqs.qza \
--o-unmatched-seqs unoriented-seqs.qza
Then run the classifier on the oriented-seqs.qza output. Or... you can do the inverse... reverse the microgreens amplicon region reference sequences instead, and then train the classifier, then classify your reads.
qiime rescript orient-seqs \
--i-sequences microgreen_id_seqs_p23SrV_f1r1.qza \
--o-oriented-seqs mg-oriented-seqs.qza \
--o-unmatched-seqs mg-unoriented-seqs.qza
qiime feature-classifier fit-classifier-naive-bayes \
--i-reference-reads mg-oriented-seqs.qza \
--i-reference-taxonomy microgreen_algaebase.qza \
--o-classifier microgreen_id_seqs_p23SrV_f1r1_oriented_classifier.qza
qiime feature-classifier classify-sklearn \
--i-classifier microgreen_id_seqs_p23SrV_f1r1_oriented_classifier.qza \
...
Another quick test could be to make use of qiime feature-classifier classify-consensus-vsearch ... as this does not care about read orientation. Although you might obtain a good taxonomy, the reads will likely be in mixed orientation and be incorrect for phylogey based diversity analyses.