There are no backward compatibility concerns I'm aware of? Can you provide an example where such code would break?

Theoretically if a 2nd argument is supplied to tostring, it will be used in this case?

local function getString(thing: T, padding: number?)
    local text = thing.name
    if padding ~= nil then
         text = string.rep(' ', padding) .. text
    end
    return text
end

setmetatable(Thing, {
    __tostring = getString,
})

-- Other calls
print(getString(thing, 4))

...would error with {thing,f}, though this isn't really backwards incompatible as it still requires you actually use the new specification.

I guess technically you could write tostring(x, y) and it'd work just fine and discards y and indeed, adding this feature would turn a dormant bug into a breaking bug. I still consider that an extremely unlikely event except in the case of variadics which makes it look inconspicuous at first. Something as weird as this most certainly would change behavior.

local t = setmetatable({ 1, 2, 3, 5 }, {
  __tostring = table.concat
})

-- before this RFC:
print(tostring(t)) -- 1235
print(tostring(t, ",")) -- 1235
print(`{t,,}`) -- illegal

-- after this RFC:
print(tostring(t)) -- 1235
print(tostring(t, ",")) -- 1,2,3,5 (behavior changed)
print(`{t,,}`) -- 1,2,3,5

But it's not backward incompatible via format specifiers in string interpolation syntax because it's syntactically invalid to even write one right now.