Fixed-point fuzzy booleans in LLVM: byte tricks that need no special cases

Storing a fuzzy truth value as a byte k meaning k/255 makes most operators fall out of existing code: fuzzy not (1 - x) is plain bitwise not on i8, since ~k = 255 - k; && and || are llvm.umin/umax. The rounded product is (a*b + 127) / 255 in i16, and since 255 is odd it never ties. Short-circuiting can share one codegen path with bool: branch on 'is the left side the absorbing value' (false/true for bool, 0/255 for the byte), and in both cases the decided result is the left side itself, so the phi takes lhs from that edge for either type. For float-to-byte conversion, a single range check with unordered compares (fcmp ult x,0 || fcmp ugt x,1) also catches NaN, so no separate NaN test is needed.