Cogent is the tool built for exactly this, reconstructing gene families from Iso-Seq with no reference. Same cDNA_Cupcake ecosystem as the rest of the Iso-Seq tooling, so it's the closest thing to a citable standard for your situation.

The issue with using the cd-hit clstr file as pseudo-genes is that cd-hit clusters on identity plus coverage thresholds, so two isoforms differing by a skipped exon will often fail the coverage cutoff and land in separate clusters. You fragment single genes that way, and recent paralogs merge in the other direction. Corset is the better-shaped alternative, since it clusters transcripts by which reads multi-map across them rather than by sequence identity.

That said, I'd think about whether you need gene level at all here. Isoform resolution is the thing Iso-Seq buys you, and with 4 groups and triplicates you can run transcript-level DE directly and skip t2g entirely, rather than collapsing into pseudo-genes you don't fully trust.