“A usable isolation is not a karaoke track” is the distinction that matters. For a pitch or timing listen, the vocal still needs breaths, consonant edges, and a little of the original reverb tail; a fully dried stem can sound impressive for a few seconds and then fail the “would I approve this in a mix?” test. Scoring plosives, sustained vowels, and hi-hat leak at a few fixed timestamps is a better habit than a single full-pass impression.

The failure modes you listed are the ones that usually waste the next hour if you stay on the same slider: kick leakage that reads as chest, soft verse phrases dropped as ambience, metallic sidebands on a dense harmony stack. Changing the source — a verse loop, a chorus-only bounce, or a light high-pass so the model is not chewing on sub — is the more useful move. Mono collapse is the other check I would not skip; phasey leftovers that disappear on speakers show up immediately on a phone.

Do you keep the gentler second pass as the default send (labeled as a pitch-review isolate, not a remix vocal), and only print a drier version when someone explicitly wants a karaoke-style lead?