On the extraction you can skip the coordinate parsing for most purposes - adding sseq to your -outfmt 6 column list hands you the aligned subject sequence directly. blastdbcmd is still what you want if you need flanking sequence rather than just the aligned block.

The translate-and-reciprocal-blast plan is where I'd be careful. These are eukaryotic genomes, so your tblastn hits are exon fragments: introns break the alignment and one protein query gives you several HSPs strung along the locus. Translating the extracted genomic span translates the introns along with it, so the thing you'd be reciprocal-blasting isn't really a protein.

For that you want a spliced protein-to-genome aligner rather than BLAST. miniprot is the easy modern option, exonerate protein2genome the older one. Either turns your protein query into an actual gene model on the assembly, and that's what's worth translating and checking reciprocally. Also answers your last question - annotating the whole genome first is overkill if you only care about one protein family.