It's good for another reason too: it decouples releases from deployments. Both deployments and releases are high-risk activities. Performing them together increases the risk multiplicatively.

Separating them does also create some incidental complexity (permutations of configurations, feature flag management) but in my experience that complexity is easier to analyse and deal with. (Dealing with feature flag complexity involves retiring them promptly, differentiating between feature flags and kill switches, etc.)

Of course, TFA knows this. They've moved the goalposts by making an extreme claim ("every" change behind a flag) and then argued against it.