document why evalRandom/formatFraction use native float

These two are the only intentional exceptions to the decimal.js-only
rule in the engine; leaving that unstated made them look like bugs
during review.

src/engine/evaluator.ts

@@ -547,6 +547,9 @@ function evalCall(name: string, args: Value[], ctx: EvalCtx): Value {
   }
 }
 
+// Native float, not Decimal: this is a PRNG (mulberry32/strHash bit-twiddling
+// over 32-bit ints), not arithmetic on user quantities — decimal.js has no
+// bitwise ops and would add nothing but overhead here.
 function evalRandom(args: Value[], ctx: EvalCtx): Value {
   const seed = strHash(ctx.line.lineText) ^ (ctx.line.index * 0x9e3779b9);
   const rand = mulberry32(seed);

src/engine/formatter.ts

@@ -109,6 +109,9 @@ function withUnit(num: string, f: string): string {
 }
 
 /** Best rational approximation via continued fractions: 0.75 -> "3/4". */
+// Native float, not Decimal: the continued-fraction algorithm is tuned to
+// float epsilon by design, and this is display formatting, not arithmetic
+// carried forward into the result.
 function formatFraction(v: Decimal): string {
   const neg = v.isNegative();
   const x = v.abs().toNumber();