Your instincts are correct, but there is a little bit more to the story.
Variant calls or genotyping chip-"called" SNPs are checked for HWE because variants found WAY outside of HWE are commonly that way due to technical artifacts.
Consider a SNP that is very close, say 4bp away from another SNP. Imagine that the SNP you have is found way out of HWE, and in fact contains an excess of heterozygous calls. Imagine that you examined MAF, and also discovered that the minor allele, which normally has an allele frequency of 0.1, in fact has a MAF of 0.13 in the same ethnicity, and nearly all the difference is in HET calls.
In this case, it is possible that the chip is actually picking up signal from the SNP that is 4bp away, and calling the SNP you are interested as the minor allele.
Now, above I said your instincts are correct. In fact, disease-relevant SNPs are out of HWE due to selection or other factors at greater than chance rates. The problem is, they are out of HWE for technical reasons even more commonly!
As a result, the conservative approach for SNPs out of HWE, especially for SNPs WAY out of HWE, is throw them out. If the SNP is in fact representative of a true finding, then other SNPs in the area should be associated with the condition as well.
Disclaimer: what I mentioned above is one very specific example that happened to me. There are MANY other examples that could be mentioned. The general take home is, SNPs out of HWE are suspicious for being technical issues, but in doing so we realize we also may be throwing out exactly the type of SNPs we want to find.
If you do in fact have a SNP that is of intense interest to you, it is probably wise to validate it in other ways, probably first bioinformatically, and second with a new assay, before spending money on functional follow up. For instance, as a quick check, does MAF seem roughly in line with what is expected for that ethnicity? Does the locus have support from LD?