Thanks @jjgao. NM_018263.4 has 1407aa while NM_018263.5 has 1435aa. Since DMP is using the older one, that's what we are sticking with in isoform_overrides_at_mskcc. Lemme know what led you to this, and we can find an alternative solution.
(Update 8/7/2017 - The info above on aa lengths is incorrect.)

I'm gonna close this request. It exposed a larger problem that versioning of RefSeq isoforms needs to be implemented across the CMO, before we do another sync up.
(Update 8/7/2017 - This comment is also incorrect. See below for details on what really went wrong.)

Per my work notes matching ENST IDs to Refseq IDs used by MSKCC's clinical bioinformatics, an ENST ID for NM_018263.4 couldn't be automatically extracted via NCBI's CCDS. So I had manually looked it up in Ensembl's release 75 archives (the latest for GRCh37 loci, viewable at feb2014.archive.ensembl.org), where the xref_refseq mapping was incorrect, as JJ pointed out in the first comment. I went back and reviewed the 8 other genes for which I had manually looked up matching ENST isoforms, and found similar mistakes in 3 other genes. Updated those too. Will merge shortly.