This workflow is inherently unsafe, and can not be made safe.

Even if we did implement partial-commit, we can't run local unit tests on only some of your changes, because there's no way to tell PHP or Python or Node or make to only look at some of the changes. Your test results and manual testing may be incorrect because of additional changes you've made locally.

You can isolate changes completely by using multiple working copies, or using a VCS with local branching (like Git or Mercurial), or by using a Git-SVN bridge to use another VCS frontend. Of these approaches, simulating local branching by using multiple working copies is likely easiest (I used multiple working copies for the duration of my time at Facebook without issue).

We do make an effort to detect when the files differ, because this is straightforward, but we could only make this inherently unsafe workflow slightly safer by looking at changes within files, and doing so would be much more complex and inherently error-prone (we can't, in the general case, tell when a change is a typo fix vs a response to feedback vs an update from upstream vs a distinct change in the same working copy).