It hinges on a single question: does the thin page own a distinct intent, or just echo a stronger page’s job? That single test routes it to one of two branches, and the word count of the page has nothing to do with which.
Expand the page when it targets a real, separate intent that deserves its own URL, a different question, a different stage of the journey, a different audience, even if it shares keywords with other pages. A page that is thin but distinct is not a duplicate; it is an underbuilt page doing a job nothing else on the site does, and the fix is to build it out until it fully answers that intent. Merge the page when it does the same job as a stronger sibling, answering the same question for the same reader, only less well. There, expanding is the wrong move, because you would be growing a page that should have been absorbed, and two pages chasing one intent split each other’s ranking signals and leave both weaker. Fold its useful parts into the stronger page and redirect it so the equity follows.
The deciding question is plain. Describe each page in one sentence. Do the two sentences describe two different reader needs, or the same need twice. Two needs means expand and keep both distinct. The same need twice means merge into the stronger URL.
The trap to avoid is the reflex to just add words. Padding a thin page does nothing if its job already belongs to a better page; you end up with two mediocre pages instead of one strong one. Put each thin page into the expand pile or the merge pile based on that one-sentence test, and don’t leave any sitting in a maybe pile, because the maybe pile is where cannibalization quietly grows.