MW core has a huge list of globally-suppressed phan issues. This weakens phan's analysis a lot, and can (and will!) cause actual errors to pass CI unnoticed. For a real, recent example, see T231488.
Running phan with all those issues suppressed exposes us to an increased risk of regressions; and on untested code, this could mean new production errors and/or train blockers.
As such, the proposal is to kill the list of suppressions (or at least, limit it to max 5 issue types), so that we can fully benefit from phan.
False positives will happen, of course. But in such a case, it's highly preferred to just suppress the issue inline (with @phan-suppress-next-line and similar), at block-level (via @suppress), or file-level (with @phan-file-suppress), so that the issue won't be suppressed in the remaining 99.9% of the code. Keeping mediawiki-phan-config up-to-date with phan should also be done in tandem: it'll be able to detect new issue types and reduce the amount of false positives.