Reports

Commentators explained that the reason why nWrap can't produce an optimized return value here is because it would change the return type of the (non-templated) function.

With that, I found that embedding nCalls return value into another lambda definition (and thus a distinct type), will enable compilers to eliminate the loop:

static inline constexpr auto nWrap(std::size_t n) {
    return [n]() { for (std::size_t i = 0; i != n; ++i) call(); };
};

std::function<void()> nCalls(std::size_t n) {
    if (n == 1) {
        return []{nWrap(1)();};  // Return value embedded into another lambda
    }
    return nWrap(n);
}

std::function<void()> oneCall() { return nCalls(1); }

Results in oneCall always returning an optimized std::function and nCalls deciding at runtime what to return – which isn't what I asked for, but it's what I'm currently settling with.

79370743