Per-sample isn't a workaround here, it's the correct thing to do. The soup is a property of each individual emulsion, so sample A's ambient profile has nothing to do with sample B's - estimating one contamination distribution across 43 channels that were never mixed would be wrong even if it fitted in memory.

Two things worth knowing. decontX takes a batch argument and does that splitting internally, so you may not need to loop by hand. And it takes z for cell population labels - pass your existing integrated annotations in there. That gives you per-sample decontamination against a consistent global label set, which is exactly the "keep my annotations" behaviour you're after. Looping per sample with per-sample clustering instead would leave you with inconsistent cluster definitions between samples and therefore an inconsistent correction.

Then yes, re-normalise and re-integrate from the corrected counts. One caution: decontXcounts come out non-integer, so if you're doing DE downstream use the raw counts for that and treat the decontamination as affecting clustering and annotation only.