@noble/post-quantum 0.6.1 → 0.7.1

package/slh-dsa.js

@@ -28,10 +28,17 @@
  */
 /*! noble-post-quantum - MIT License (c) 2024 Paul Miller (paulmillr.com) */
 import { hmac } from '@noble/hashes/hmac.js';
+import { bytesToNumberBE, numberToBytesBE } from '@noble/curves/utils.js';
 import { sha256, sha512 } from '@noble/hashes/sha2.js';
 import { shake256 } from '@noble/hashes/sha3.js';
-import { bytesToHex, concatBytes, createView, hexToBytes, } from '@noble/hashes/utils.js';
+import { concatBytes, createView } from '@noble/hashes/utils.js';
 import { abytes, checkHash, cleanBytes, copyBytes, equalBytes, getMask, getMessage, getMessagePrehash, randomBytes, splitCoder, validateSigOpts, validateVerOpts, vecCoder, } from "./utils.js";
+// Keys the internal SLH-DSA surface accepts. `context` is deliberately absent: the public
+// wrappers consume it when they format M' and must not forward it, because a key that is
+// accepted and then never read is the same silent downgrade this validation exists to prevent.
+// `extraEntropy` is signing-only, so verification (which takes no options of its own) has none.
+const INTERNAL_SIG_OPT_KEYS = /* @__PURE__ */ Object.freeze(['extraEntropy']);
+const INTERNAL_VER_OPT_KEYS = /* @__PURE__ */ Object.freeze([]);
 /** Winternitz signature params. */
 /**
  * Built-in SLH-DSA Table 2 subset keyed by strength/profile.
@@ -61,19 +68,6 @@ const AddressType = {
     WOTSPRF: 5,
     FORSPRF: 6,
 };
-function hexToNumber(hex) {
-    if (typeof hex !== 'string')
-        throw new Error('hex string expected, got ' + typeof hex);
-    return BigInt(hex === '' ? '0' : '0x' + hex); // Big Endian
-}
-// BE: Big Endian, LE: Little Endian. This is the local FIPS 205 `toInt(...)` equivalent.
-function bytesToNumberBE(bytes) {
-    return hexToNumber(bytesToHex(bytes));
-}
-// Local in-range FIPS 205 `toByte(x, n)` equivalent; callers must keep `n < 256^len`.
-function numberToBytesBE(n, len) {
-    return hexToBytes(n.toString(16).padStart(len * 2, '0'));
-}
 // Local FIPS 205 Algorithm 4 `base_2^b(...)` implementation. Bits are consumed in big-endian
 // order within each input byte, and callers must provide at least `ceil(outLen * b / 8)` bytes;
 // short inputs are not rejected and would zero-extend implicitly.
@@ -92,8 +86,11 @@ const base2b = (outLen, b) => {
         return baseB;
     };
 };
+const _1n = /* @__PURE__ */ BigInt(1);
+const _8n = /* @__PURE__ */ BigInt(8);
+const _0xffn = /* @__PURE__ */ BigInt(0xff);
 function getMaskBig(bits) {
-    return (1n << BigInt(bits)) - 1n; // 4 -> 0b1111
+    return (_1n << BigInt(bits)) - _1n; // 4 -> 0b1111
 }
 /** One parameter/hash instantiation of the public SLH-DSA API.
  * `keygen(seed)` is a deterministic 3N-byte library hook around the internal keygen flow,
@@ -137,9 +134,18 @@ function gen(opts, hashOpts_) {
     // `height` / `chain` and `index` / `hash` share the same spec words, so callers must use the
     // address-type-specific combinations instead of mixing both meanings in one call.
     const setAddr = (opts, addr = new Uint8Array(ADDR_BYTES)) => {
-        const { type, height, tree, layer, index, chain, hash, keypair } = opts;
-        const { subtreeAddr, keypairAddr } = opts;
-        const v = createView(addr);
+        // These objects are created in hot internal loops, so avoid cloning them. Read only own fields:
+        // absent address words must stay absent even if Object.prototype was polluted.
+        const type = Object.hasOwn(opts, 'type') ? opts.type : undefined;
+        const height = Object.hasOwn(opts, 'height') ? opts.height : undefined;
+        const tree = Object.hasOwn(opts, 'tree') ? opts.tree : undefined;
+        const layer = Object.hasOwn(opts, 'layer') ? opts.layer : undefined;
+        const index = Object.hasOwn(opts, 'index') ? opts.index : undefined;
+        const chain = Object.hasOwn(opts, 'chain') ? opts.chain : undefined;
+        const hash = Object.hasOwn(opts, 'hash') ? opts.hash : undefined;
+        const keypair = Object.hasOwn(opts, 'keypair') ? opts.keypair : undefined;
+        const subtreeAddr = Object.hasOwn(opts, 'subtreeAddr') ? opts.subtreeAddr : undefined;
+        const keypairAddr = Object.hasOwn(opts, 'keypairAddr') ? opts.keypairAddr : undefined;
         if (height !== undefined)
             addr[OFFSET_CHAIN_ADDR] = height;
         if (layer !== undefined)
@@ -150,12 +156,21 @@ function gen(opts, hashOpts_) {
             addr[OFFSET_CHAIN_ADDR] = chain;
         if (hash !== undefined)
             addr[OFFSET_HASH_ADDR] = hash;
-        if (index !== undefined)
-            v.setUint32(OFFSET_TREE_INDEX, index, false);
+        // Manual big-endian writes: setAddr runs in the innermost WOTS/tree loops, and creating a
+        // DataView per call was a measurable share of sign() time.
+        if (index !== undefined) {
+            addr[OFFSET_TREE_INDEX + 0] = index >>> 24;
+            addr[OFFSET_TREE_INDEX + 1] = index >>> 16;
+            addr[OFFSET_TREE_INDEX + 2] = index >>> 8;
+            addr[OFFSET_TREE_INDEX + 3] = index;
+        }
         if (subtreeAddr)
             addr.set(subtreeAddr.subarray(0, OFFSET_TREE + 8));
-        if (tree !== undefined)
-            v.setBigUint64(OFFSET_TREE, tree, false);
+        if (tree !== undefined) {
+            let t = tree;
+            for (let i = 7; i >= 0; i--, t >>= _8n)
+                addr[OFFSET_TREE + i] = Number(t & _0xffn);
+        }
         if (keypair !== undefined) {
             addr[OFFSET_KP_ADDR1] = keypair;
             if (TREE_HEIGHT > 8)
@@ -209,10 +224,12 @@ function gen(opts, hashOpts_) {
         const maxIdx = (1 << height) - 1;
         const stack = new Uint8Array(height * N);
         const authPath = new Uint8Array(height * N);
+        // One node buffer per treehash call (not per leaf): both halves are fully overwritten at
+        // each use, and the returned root aliases cur1, which is never reused after return.
+        const current = new Uint8Array(2 * N);
+        const cur0 = current.subarray(0, N);
+        const cur1 = current.subarray(N);
         for (let idx = 0;; idx++) {
-            const current = new Uint8Array(2 * N);
-            const cur0 = current.subarray(0, N);
-            const cur1 = current.subarray(N);
             const addrOffset = idx + idxOffset;
             cur1.set(leafFn(leafIdx, addrOffset, rawContext, info));
             let h = 0;
@@ -360,7 +377,7 @@ function gen(opts, hashOpts_) {
             return Uint8Array.from(pk);
         },
         sign: (msg, sk, opts = {}) => {
-            validateSigOpts(opts);
+            opts = validateSigOpts(opts, INTERNAL_SIG_OPT_KEYS);
             let { extraEntropy: random } = opts;
             const [skSeed, skPRF, pk] = secretCoder.decode(sk); // todo: fix
             const [pkSeed, _] = publicCoder.decode(pk);
@@ -394,7 +411,9 @@ function gen(opts, hashOpts_) {
                     height: 0,
                     index: indices[i] + idxOffset,
                 }, forsTreeAddr);
-                const prf = context.PRFaddr(forsTreeAddr);
+                // Copy: PRFaddr returns a per-context scratch view, and this value is retained in
+                // `fors` across the many PRFaddr calls inside forsTreehash below.
+                const prf = copyBytes(context.PRFaddr(forsTreeAddr));
                 setAddr({ type: AddressType.FORSTREE }, forsTreeAddr);
                 const { root, authPath } = forsTreehash(context, indices[i], idxOffset, forsTreeAddr, forsLeaf);
                 roots.push(root);
@@ -404,7 +423,9 @@ function gen(opts, hashOpts_) {
                 type: AddressType.FORSPK,
                 keypairAddr: wotsAddr,
             });
-            const root = context.thashN(K, concatBytes(...roots), forsPkAddr);
+            // Copy: thashN returns a per-context scratch view, and `root` lives across every hash
+            // call in the hypertree loop below (it is also mutated via root.set).
+            const root = copyBytes(context.thashN(K, concatBytes(...roots), forsPkAddr));
             // WOTS signatures
             const treeAddr = setAddr({ type: AddressType.HASHTREE });
             const wots = [];
@@ -422,12 +443,20 @@ function gen(opts, hashOpts_) {
             cleanBytes(R, random, treeAddr, wotsAddr, forsLeaf, forsTreeAddr, indices, roots);
             return SIG;
         },
-        verify: (sig, msg, publicKey) => {
+        verify: (sig, msg, publicKey, opts = {}) => {
+            // The internal verify reads no options; reject any so a stray key (e.g. a caller
+            // mistaking this for the public verify and passing `context`) is reported rather than
+            // silently swallowed by this function's arity.
+            validateVerOpts(opts, INTERNAL_VER_OPT_KEYS);
             const [pkSeed, pubRoot] = publicCoder.decode(publicKey);
-            const [random, forsVec, wotsVec] = sigCoder.decode(sig);
             const pk = publicKey;
+            // FIPS 205 Algorithm 20 step 1: wrong-length signatures return false instead of throwing
+            // (same as ml-dsa). Must run before sigCoder.decode, which throws on length mismatch.
+            // Preserve TypeError for non-byte API arguments before treating byte lengths as invalid.
+            abytes(sig, undefined, 'signature');
             if (sig.length !== sigCoder.bytesLen)
                 return false;
+            const [random, forsVec, wotsVec] = sigCoder.decode(sig);
             const context = getContext(pkSeed);
             let { tree, leafIdx, md } = hashMessage(random, pk, msg, context);
             const wotsAddr = setAddr({
@@ -447,14 +476,16 @@ function gen(opts, hashOpts_) {
                 const idxOffset = i << A;
                 setAddr({ height: 0, index: indices[i] + idxOffset }, forsTreeAddr);
                 const leaf = context.thash1(prf, forsTreeAddr);
-                // Compute inplace, because we need all roots in same byte array
-                roots.push(computeRoot(leaf, indices[i], idxOffset, authPath, A, context, forsTreeAddr));
+                // Copy: computeRoot returns a thashN scratch view, and roots are retained across the
+                // remaining FORS iterations (computeRoot itself copies `leaf` before hashing).
+                roots.push(copyBytes(computeRoot(leaf, indices[i], idxOffset, authPath, A, context, forsTreeAddr)));
             }
             const forsPkAddr = setAddr({
                 type: AddressType.FORSPK,
                 keypairAddr: wotsAddr,
             });
-            let root = context.thashN(K, concatBytes(...roots), forsPkAddr); // root = thash()
+            // Copy: `root` must survive the thash1/thashN calls of the WOTS chain loop below.
+            let root = copyBytes(context.thashN(K, concatBytes(...roots), forsPkAddr)); // root = thash()
             // WOTS signature
             const treeAddr = setAddr({ type: AddressType.HASHTREE });
             const wotsPkAddr = setAddr({ type: AddressType.WOTSPK });
@@ -477,7 +508,8 @@ function gen(opts, hashOpts_) {
                     }
                 }
                 const leaf = context.thashN(WOTS_LEN, wotsPk, wotsPkAddr);
-                root = computeRoot(leaf, leafIdx, 0, sigAuth, TREE_HEIGHT, context, treeAddr);
+                // Copy: `root` is read by chainLengths / equalBytes after later hash calls.
+                root = copyBytes(computeRoot(leaf, leafIdx, 0, sigAuth, TREE_HEIGHT, context, treeAddr));
                 leafIdx = Number(tree & getMaskBig(TREE_HEIGHT));
             }
             return equalBytes(root, pubRoot);
@@ -491,14 +523,16 @@ function gen(opts, hashOpts_) {
         keygen: internal.keygen,
         getPublicKey: internal.getPublicKey,
         sign: (msg, secretKey, opts = {}) => {
-            validateSigOpts(opts);
+            opts = validateSigOpts(opts);
             const M = getMessage(msg, opts.context);
-            const res = internal.sign(M, secretKey, opts);
+            // `context` is consumed by getMessage() above; forwarding it would make the internal
+            // surface accept a key it never reads.
+            const res = internal.sign(M, secretKey, { extraEntropy: opts.extraEntropy });
             cleanBytes(M);
             return res;
         },
         verify: (sig, msg, publicKey, opts = {}) => {
-            validateVerOpts(opts);
+            opts = validateVerOpts(opts);
             return internal.verify(sig, getMessage(msg, opts.context), publicKey);
         },
         prehash: (hash) => {
@@ -510,14 +544,15 @@ function gen(opts, hashOpts_) {
                 keygen: internal.keygen,
                 getPublicKey: internal.getPublicKey,
                 sign: (msg, secretKey, opts = {}) => {
-                    validateSigOpts(opts);
+                    opts = validateSigOpts(opts);
                     const M = getMessagePrehash(rawHash, msg, opts.context);
-                    const res = internal.sign(M, secretKey, opts);
+                    // As above: getMessagePrehash() consumes `context`, so it must not travel further.
+                    const res = internal.sign(M, secretKey, { extraEntropy: opts.extraEntropy });
                     cleanBytes(M);
                     return res;
                 },
                 verify: (sig, msg, publicKey, opts = {}) => {
-                    validateVerOpts(opts);
+                    opts = validateVerOpts(opts);
                     return internal.verify(sig, getMessagePrehash(rawHash, msg, opts.context), publicKey);
                 },
             });
@@ -533,21 +568,26 @@ const genShake = () => (opts) => (pubSeed, skSeed) => {
     // for each address-bound call instead of reabsorbing the same seed every time.
     const h0 = shake256.create({}).update(pubSeed);
     const h0tmp = h0.clone();
+    // Per-context output scratch: thash1/thashN/PRFaddr return these buffers directly, so
+    // callers must consume or copy a result before the next call on the same lane.
+    const thashOut = new Uint8Array(N);
+    const prfOut = new Uint8Array(N);
     const thash = (blocks, input, addr) => {
         stats.thash++;
-        return h0
-            ._cloneInto(h0tmp)
+        const len = blocks * N;
+        h0._cloneInto(h0tmp)
             .update(addr)
-            .update(input.subarray(0, blocks * N))
-            .xof(N);
+            .update(input.length === len ? input : input.subarray(0, len))
+            .xofInto(thashOut);
+        return thashOut;
     };
     return {
         PRFaddr: (addr) => {
             if (!skSeed)
                 throw new Error('no sk seed');
             stats.prf++;
-            const res = h0._cloneInto(h0tmp).update(addr).update(skSeed).xof(N);
-            return res;
+            h0._cloneInto(h0tmp).update(addr).update(skSeed).xofInto(prfOut);
+            return prfOut;
         },
         PRFmsg: (skPRF, random, msg) => {
             stats.gen_message_random++;
@@ -568,11 +608,12 @@ const genShake = () => (opts) => (pubSeed, skSeed) => {
         clean: () => {
             h0.destroy();
             h0tmp.destroy();
+            cleanBytes(thashOut, prfOut);
             //console.log(stats);
         },
     };
 };
-const SHAKE_SIMPLE = /* @__PURE__ */ (() => ({ getContext: genShake() }))();
+const SHAKE_SIMPLE = /* @__PURE__ */ (() => ({ isCompressed: false, getContext: genShake() }))();
 /**
  * SLH-DSA-SHAKE-128f: Table 2 row `n=16, h=66, d=22, h'=3, a=6, k=33, lg w=4, m=34`;
  * lengths `publicKey=32`, `secretKey=64`, `signature=17088`, `seed=48`, `signRand=16`.
@@ -635,6 +676,16 @@ const genSha = (h0, h1) => (opts) => (pub_seed, sk_seed) => {
         .update(new Uint8Array(h1.blockLen - N));
     const h0tmp = h0ps.clone();
     const h1tmp = h1ps.clone();
+    // Per-context output scratch: thash1/thashN/PRFaddr return views into these buffers, so
+    // callers must consume or copy a result before the next call on the same lane (see Context
+    // docs). digestInto also skips digest()'s per-call destroy(): the tmp states are fully
+    // overwritten by the next _cloneInto and wiped in clean().
+    const h0out = new Uint8Array(h0.outputLen);
+    const h1out = new Uint8Array(h1.outputLen);
+    const prfOut = new Uint8Array(h0.outputLen);
+    const h0outN = h0out.subarray(0, N);
+    const h1outN = h1out.subarray(0, N);
+    const prfOutN = prfOut.subarray(0, N);
     // https://www.rfc-editor.org/rfc/rfc8017.html#appendix-B.2.1
     // This local helper is intentionally stricter than generic MGF1 reuse: current SLH-DSA callers
     // only request tiny `m`-byte outputs, but the guard below rejects `length > 2^32` instead of
@@ -653,27 +704,26 @@ const genSha = (h0, h1) => (opts) => (pub_seed, sk_seed) => {
         cleanBytes(out.subarray(length));
         return out.subarray(0, length);
     }
-    const thash = (_, h, hTmp) => (blocks, input, addr) => {
+    const thash = (h, hTmp, out, outN) => (blocks, input, addr) => {
         stats.thash++;
-        const d = h
-            ._cloneInto(hTmp)
+        const len = blocks * N;
+        h._cloneInto(hTmp)
             .update(addr)
-            .update(input.subarray(0, blocks * N))
-            .digest();
-        return d.subarray(0, N);
+            .update(input.length === len ? input : input.subarray(0, len))
+            .digestInto(out);
+        return outN;
     };
     return {
         PRFaddr: (addr) => {
             if (!sk_seed)
                 throw new Error('No sk seed');
             stats.prf++;
-            const res = h0ps
+            h0ps
                 ._cloneInto(h0tmp)
                 .update(addr)
                 .update(sk_seed)
-                .digest()
-                .subarray(0, N);
-            return res;
+                .digestInto(prfOut);
+            return prfOutN;
         },
         PRFmsg: (skPRF, random, msg) => {
             stats.gen_message_random++;
@@ -689,13 +739,14 @@ const genSha = (h0, h1) => (opts) => (pub_seed, sk_seed) => {
             const seed = concatBytes(R.subarray(0, N), pk.subarray(0, N), h1.create().update(R.subarray(0, N)).update(pk).update(m).digest());
             return mgf1(seed, outLen, h1);
         },
-        thash1: thash(h0, h0ps, h0tmp).bind(null, 1),
-        thashN: thash(h1, h1ps, h1tmp),
+        thash1: thash(h0ps, h0tmp, h0out, h0outN).bind(null, 1),
+        thashN: thash(h1ps, h1tmp, h1out, h1outN),
         clean: () => {
             h0ps.destroy();
             h1ps.destroy();
             h0tmp.destroy();
             h1tmp.destroy();
+            cleanBytes(h0out, h1out, prfOut);
             //console.log(stats);
         },
     };
@@ -712,6 +763,23 @@ const SHA512_SIMPLE = /* @__PURE__ */ (() => ({
  * SLH-DSA-SHA2-128f: Table 2 row `n=16, h=66, d=22, h'=3, a=6, k=33, lg w=4, m=34`;
  * lengths `publicKey=32`, `secretKey=64`, `signature=17088`, `seed=48`, `signRand=16`.
  * Also exposes `.prehash(...)`.
+ * @example
+ * Generate deterministic SLH-DSA keys, sign one message, and verify the signature.
+ * ```ts
+ * import { sha256 } from '@noble/hashes/sha2.js';
+ * import { slh_dsa_sha2_128f } from '@noble/post-quantum/slh-dsa.js';
+ * const seed = new Uint8Array(slh_dsa_sha2_128f.lengths.seed!);
+ * const { secretKey, publicKey } = slh_dsa_sha2_128f.keygen(seed);
+ * const msg = new TextEncoder().encode('hello noble');
+ * const sig = slh_dsa_sha2_128f.sign(msg, secretKey);
+ * const isValid = slh_dsa_sha2_128f.verify(sig, msg, publicKey);
+ * const recovered = slh_dsa_sha2_128f.getPublicKey(secretKey);
+ * const context = new Uint8Array([1, 2, 3]);
+ * const prehash = slh_dsa_sha2_128f.prehash(sha256);
+ * const preSig = prehash.sign(msg, secretKey, { context });
+ * const preValid = prehash.verify(preSig, msg, publicKey, { context });
+ * const internalSig = slh_dsa_sha2_128f.internal.sign(msg, secretKey);
+ * ```
  */
 export const slh_dsa_sha2_128f = /* @__PURE__ */ (() => gen(PARAMS['128f'], SHA256_SIMPLE))();
 /**
@@ -744,4 +812,3 @@ export const slh_dsa_sha2_256f = /* @__PURE__ */ (() => gen(PARAMS['256f'], SHA5
  * Also exposes `.prehash(...)`.
  */
 export const slh_dsa_sha2_256s = /* @__PURE__ */ (() => gen(PARAMS['256s'], SHA512_SIMPLE))();
-//# sourceMappingURL=slh-dsa.js.map

package/src/_crystals.ts

@@ -73,9 +73,15 @@ type Crystals<T extends TypedArray> = {
   smod: (a: number, modulo?: number) => number;
   nttZetas: T;
   NTT: {
-    /** Forward transform in place. Mutates and returns `r`. */
+    /**
+     * Forward transform in place. Mutates and returns `r`.
+     * Kyber-mode input coefficients must already use canonical representatives in `[0, Q)`.
+     */
     encode: (r: T) => T;
-    /** Inverse transform in place. Mutates and returns `r`. */
+    /**
+     * Inverse transform in place. Mutates and returns `r`.
+     * Kyber-mode input coefficients must already use canonical representatives in `[0, Q)`.
+     */
     decode: (r: T) => T;
   };
   bitsCoder: (d: number, c: Coder<number, number>) => BytesCoderLen<T>;
@@ -106,7 +112,7 @@ export const genCrystals = <T extends TypedArray>(opts: CrystalOpts<T>): TRet<Cr
   // Normalize JS `%` into the canonical Z_m representative `[0, modulo-1]` expected by
   // FIPS 203 §2.3 / FIPS 204 §2.3 before downstream mod-q arithmetic.
   const mod = (a: number, modulo = Q): number => {
-    const result = a % modulo | 0;
+    const result = (a % modulo) | 0;
     return (result >= 0 ? result | 0 : (modulo + result) | 0) | 0;
   };
   // FIPS 204 §7.4 uses the centered `mod ±` representative for low bits, keeping the
@@ -135,14 +141,34 @@ export const genCrystals = <T extends TypedArray>(opts: CrystalOpts<T>): TRet<Cr
   // Kyber has slightly different params, since there is no 512th primitive root of unity mod q,
   // only 256th primitive root of unity mod. Which also complicates MultiplyNTT.
 
-  const field = {
-    add: (a: number, b: number) => mod((a | 0) + (b | 0)) | 0,
-    sub: (a: number, b: number) => mod((a | 0) - (b | 0)) | 0,
-    mul: (a: number, b: number) => mod((a | 0) * (b | 0)) | 0,
-    inv: (_a: number) => {
-      throw new Error('not implemented');
-    },
+  const inv = (_a: number) => {
+    throw new Error('not implemented');
   };
+  // ML-KEM (Kyber) polynomials always enter the transform reduced to [0, Q), so add/sub only
+  // need one conditional correction instead of `%`; measured ~20% faster NTT there.
+  // ML-DSA keeps the generic mod() path on purpose: its first forward stage sees centered
+  // (negative) coefficients, and `sub(a, t)` can drop below -Q (t is a mul output in [0, Q)),
+  // so a single correction is not enough. A guarded fast path with mod() fallback was measured
+  // slower than plain `%` for the 23-bit Q (V8 int32 modulo is one div; the branches lose).
+  const field = isKyber
+    ? {
+        add: (a: number, b: number) => {
+          const r = (a + b) | 0;
+          return r >= Q ? (r - Q) | 0 : r;
+        },
+        sub: (a: number, b: number) => {
+          const r = (a - b) | 0;
+          return r < 0 ? (r + Q) | 0 : r;
+        },
+        mul: (a: number, b: number) => mod((a | 0) * (b | 0)) | 0,
+        inv,
+      }
+    : {
+        add: (a: number, b: number) => mod((a | 0) + (b | 0)) | 0,
+        sub: (a: number, b: number) => mod((a | 0) - (b | 0)) | 0,
+        mul: (a: number, b: number) => mod((a | 0) * (b | 0)) | 0,
+        inv,
+      };
   const nttOpts = {
     N,
     roots: nttZetas as any,
@@ -168,6 +194,12 @@ export const genCrystals = <T extends TypedArray>(opts: CrystalOpts<T>): TRet<Cr
   // Pack one little-endian `d`-bit word per coefficient, matching FIPS 203 ByteEncode /
   // ByteDecode and the FIPS 204 BitsToBytes-based polynomial packing helpers.
   const bitsCoder = (d: number, c: Coder<number, number>): TRet<BytesCoderLen<T>> => {
+    // Validate the carry shape once: JS bitwise operations silently truncate wider accumulators.
+    for (let i = 0, bufLen = 0; i < N; i++) {
+      bufLen += d;
+      if (bufLen > 32) getMask(bufLen);
+      bufLen %= 8;
+    }
     const mask = getMask(d);
     const bytesLen = d * (N / 8);
     return {
@@ -178,7 +210,9 @@ export const genCrystals = <T extends TypedArray>(opts: CrystalOpts<T>): TRet<Cr
         for (let i = 0, buf = 0, bufLen = 0, pos = 0; i < poly.length; i++) {
           buf |= (c.encode(poly[i]) & mask) << bufLen;
           bufLen += d;
-          for (; bufLen >= 8; bufLen -= 8, buf >>= 8) r[pos++] = buf & getMask(bufLen);
+          // Take the low byte directly: `& 0xff` matches the previous getMask(bufLen) result
+          // after Uint8Array truncation, without a validated function call per output byte.
+          for (; bufLen >= 8; bufLen -= 8, buf >>= 8) r[pos++] = buf & 0xff;
         }
         return r as TRet<Uint8Array>;
       },