numeric-leaves vectors gain two float32 rows; float32 integer rounding drops its redundant guard

An odd kept value just past exact pins the tie rule (the Stryker survivor). The excess <= 0 guard was equivalent to the
general path for the only values that reach it, so it is deleted.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

packages/cultcache-ts/src/codec.ts

@@ -88,11 +88,11 @@ export const int32 = integerCodec("System.Int32", "an int32", -(2n ** 31n), 2n *
 export const int64 = integerCodec("System.Int64", "an int64", -(2n ** 63n), 2n ** 63n - 1n);
 
 // A 64-bit integer to the nearest float32 in one rounding, ties to even, as the reference's cast does. Number(value)
-// first would round to a double and then again, and the two roundings can disagree.
+// first would round to a double and then again, and the two roundings can disagree. A bigint reaches here only beyond
+// 2^53 (the wire reader returns a number for any smaller integer), so at least 30 bits are rounded away.
 function bigintToFloat32(value: bigint): number {
   const magnitude = value < 0n ? -value : value;
   const excess = magnitude.toString(2).length - 24;
-  if (excess <= 0) return Number(value);
   const kept = magnitude >> BigInt(excess);
   const rest = magnitude - (kept << BigInt(excess));
   const half = 1n << BigInt(excess - 1);

packages/cultcache-ts/test/union-parity.test.ts

@@ -548,7 +548,7 @@ const isNaNBits = (bitsHex: string) => {
 
 test("numeric leaves read and refuse exactly what the C# reference does", async () => {
   const rows = readWire(new Uint8Array(await readFile(join(numericVectors, "reads.msgpack")))) as [string, Uint8Array, boolean, Uint8Array][];
-  assert.equal(rows.length, 148);
+  assert.equal(rows.length, 156);
   let accepted = 0;
   let refused = 0;
   for (const [leaf, payload, wasAccepted, bits] of rows) {