Yes, but even if your language is total you have an analogue of the halting problem, where instead of analysis being undecidable it just takes an unrealistic amount of effort.
My understanding is more hand-wavy than formal, but would the combination of a total language and some limitation on the total number of instructions produce a provably predictable language?
Likely not. It takes only a very small number of instructions to implement an interpreter for a Turning-complete language, and then you can create arbitrary behavior by passing different input. The number of necessary instructions is shockingly small, if using something like Binary Lambda Calculus or Subtract-And-Branch-If-Negative.