I'd do this in two steps. 1, as above, use samtools with a bed filter to get the reads where read2 do map to the region you want.

Then pull out the read names of those reads. Then use picard filterSamReads to get the read1 and read2 data belonging to those reads.

I'd definitely do this with pysam. What issues are you having? I'd do something like:

import pysam
inbam = pysam.AlignmentFile("my.bam")
outbam = pysam.AlignmentFile("my.filtered.bam", "w", template=inbam)

for read in inbam.fetch("sequence_name"):

    if read.reference_end >= 440:
        outbam.write(read)

outbam.close()