mldsa_gh 0.1.0

data/ext/mldsa_gh_native/vendor/mldsa-native/mldsa/src/sign.c

@@ -0,0 +1,1720 @@
+/*
+ * Copyright (c) The mldsa-native project authors
+ * SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT
+ */
+
+/* References
+ * ==========
+ *
+ * - [FIPS140_3_IG]
+ *   Implementation Guidance for FIPS 140-3 and the Cryptographic Module
+ *   Validation Program
+ *   National Institute of Standards and Technology
+ *   https://csrc.nist.gov/projects/cryptographic-module-validation-program/fips-140-3-ig-announcements
+ *
+ * - [FIPS204]
+ *   FIPS 204 Module-Lattice-Based Digital Signature Standard
+ *   National Institute of Standards and Technology
+ *   https://csrc.nist.gov/pubs/fips/204/final
+ *
+ * - [FIPS204_UPDATES]
+ *   FIPS 204 Potential Updates (Errata)
+ *   National Institute of Standards and Technology
+ *   https://csrc.nist.gov/files/pubs/fips/204/final/docs/fips-204-potential-updates.xlsx
+ *
+ * - [Round3_Spec]
+ *   CRYSTALS-Dilithium Algorithm Specifications and Supporting Documentation
+ *   (Version 3.1)
+ *   Bai, Ducas, Kiltz, Lepoint, Lyubashevsky, Schwabe, Seiler, Stehlé
+ *   https://pq-crystals.org/dilithium/data/dilithium-specification-round3-20210208.pdf
+ */
+
+#include "sign.h"
+
+#include "cbmc.h"
+#include "ct.h"
+#include "debug.h"
+#include "packing.h"
+#include "poly.h"
+#include "poly_kl.h"
+#include "polyvec.h"
+#include "randombytes.h"
+#include "symmetric.h"
+
+/* Parameter set namespacing
+ * This is to facilitate building multiple instances
+ * of mldsa-native (e.g. with varying parameter sets)
+ * within a single compilation unit. */
+#define mld_check_pct MLD_ADD_PARAM_SET(mld_check_pct) MLD_CONTEXT_PARAMETERS_2
+#define mld_sample_s1_s2 MLD_ADD_PARAM_SET(mld_sample_s1_s2)
+#define mld_validate_hash_length MLD_ADD_PARAM_SET(mld_validate_hash_length)
+#define mld_get_hash_oid MLD_ADD_PARAM_SET(mld_get_hash_oid)
+#define mld_H MLD_ADD_PARAM_SET(mld_H)
+#define mld_compute_pack_z MLD_ADD_PARAM_SET(mld_compute_pack_z)
+#define mld_attempt_signature_generation \
+  MLD_ADD_PARAM_SET(mld_attempt_signature_generation) MLD_CONTEXT_PARAMETERS_8
+#define mld_compute_pack_t0_t1 \
+  MLD_ADD_PARAM_SET(mld_compute_pack_t0_t1) MLD_CONTEXT_PARAMETERS_5
+#define mld_get_max_signing_attempts \
+  MLD_ADD_PARAM_SET(mld_get_max_signing_attempts)
+
+#if !defined(MLD_CONFIG_NO_KEYPAIR_API)
+static int mld_check_pct(uint8_t const pk[MLDSA_CRYPTO_PUBLICKEYBYTES],
+                         uint8_t const sk[MLDSA_CRYPTO_SECRETKEYBYTES],
+                         MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+__contract__(
+  requires(memory_no_alias(pk, MLDSA_CRYPTO_PUBLICKEYBYTES))
+  requires(memory_no_alias(sk, MLDSA_CRYPTO_SECRETKEYBYTES))
+  ensures(return_value == 0 || return_value == MLD_ERR_OUT_OF_MEMORY ||
+          return_value == MLD_ERR_RNG_FAIL ||
+          return_value == MLD_ERR_SIGNING_PAUSED ||
+          return_value == MLD_ERR_PCT_FAIL)
+);
+
+#if defined(MLD_CONFIG_KEYGEN_PCT)
+/**
+ * Pair-wise Consistency Test (PCT) for DSA keypairs.
+ *
+ * @[FIPS140_3_IG] TE10.35.02
+ * (https://csrc.nist.gov/csrc/media/Projects/cryptographic-module-validation-program/documents/fips%20140-3/FIPS%20140-3%20IG.pdf).
+ *
+ * Validates that a generated public/private key pair can correctly sign and
+ * verify data. Performs signature generation using the private key (sk),
+ * followed by signature verification using the public key (pk).
+ *
+ * @note @[FIPS204] requires that public/private key pairs are to be used
+ * only for the calculation and/or verification of digital signatures.
+ *
+ * @param[in] pk      Public key.
+ * @param[in] sk      Secret key.
+ * @param     context Application context. Only present when
+ *                    MLD_CONFIG_CONTEXT_PARAMETER is defined; type set by
+ *                    MLD_CONFIG_CONTEXT_PARAMETER_TYPE.
+ *
+ * @retval 0                      Success.
+ * @retval MLD_ERR_OUT_OF_MEMORY  MLD_CONFIG_CUSTOM_ALLOC_FREE was used and an
+ *                                allocation via MLD_CUSTOM_ALLOC returned NULL.
+ * @retval MLD_ERR_RNG_FAIL       Random number generation failed.
+ * @retval MLD_ERR_SIGNING_PAUSED The PCT's signing step was paused by a
+ *                                MLD_CONFIG_SIGN_HOOK_ATTEMPT hook.
+ * @retval MLD_ERR_PCT_FAIL       The consistency check failed.
+ */
+static int mld_check_pct(uint8_t const pk[MLDSA_CRYPTO_PUBLICKEYBYTES],
+                         uint8_t const sk[MLDSA_CRYPTO_SECRETKEYBYTES],
+                         MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  MLD_ALIGN uint8_t message[1] = {0};
+  int ret;
+  MLD_ALLOC(signature, uint8_t, MLDSA_CRYPTO_BYTES, context);
+  MLD_ALLOC(pk_test, uint8_t, MLDSA_CRYPTO_PUBLICKEYBYTES, context);
+
+  if (signature == NULL || pk_test == NULL)
+  {
+    ret = MLD_ERR_OUT_OF_MEMORY;
+    goto cleanup;
+  }
+
+  /* Copy public key for testing */
+  mld_memcpy(pk_test, pk, MLDSA_CRYPTO_PUBLICKEYBYTES);
+
+  /* Sign a test message using the original secret key */
+  ret = mld_sign_signature(signature, message, sizeof(message), NULL, 0, sk,
+                           context);
+  if (ret != 0)
+  {
+    goto cleanup;
+  }
+
+#if defined(MLD_CONFIG_KEYGEN_PCT_BREAKAGE_TEST)
+  /* Deliberately break public key for testing purposes */
+  if (mld_break_pct())
+  {
+    pk_test[0] = ~pk_test[0];
+  }
+#endif /* MLD_CONFIG_KEYGEN_PCT_BREAKAGE_TEST */
+
+  /* Verify the signature using the (potentially corrupted) public key. */
+  ret = mld_sign_verify(signature, message, sizeof(message), NULL, 0, pk_test,
+                        context);
+  if (ret != 0)
+  {
+    goto cleanup;
+  }
+
+cleanup:
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  MLD_FREE(pk_test, uint8_t, MLDSA_CRYPTO_PUBLICKEYBYTES, context);
+  MLD_FREE(signature, uint8_t, MLDSA_CRYPTO_BYTES, context);
+
+  /* A failed signing operation or an invalid signature hint at a faulty
+   * implementation and map to a dedicated error code for PCT failure. */
+  if (ret == MLD_ERR_INVALID_SIGNATURE ||
+      ret == MLD_ERR_SIGN_ATTEMPTS_EXHAUSTED)
+  {
+    ret = MLD_ERR_PCT_FAIL;
+  }
+
+  /* Other error codes, e.g. platform failures like out of memory or
+   * randomness failure, are passed on unmodified. */
+
+  return ret;
+}
+#else  /* MLD_CONFIG_KEYGEN_PCT */
+static int mld_check_pct(uint8_t const pk[MLDSA_CRYPTO_PUBLICKEYBYTES],
+                         uint8_t const sk[MLDSA_CRYPTO_SECRETKEYBYTES],
+                         MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  /* Skip PCT */
+  ((void)pk);
+  ((void)sk);
+  MLD_CONTEXT_UNUSED(context);
+  return 0;
+}
+#endif /* !MLD_CONFIG_KEYGEN_PCT */
+
+/**
+ * Sample the short secret vectors s1 (length MLDSA_L) and s2 (length MLDSA_K)
+ * with coefficients in [-MLDSA_ETA, MLDSA_ETA] from the seed.
+ *
+ * @spec{Implements @[FIPS204, Algorithm 33, ExpandS].}
+ *
+ * @param[out] s1   Output vector s1.
+ * @param[out] s2   Output vector s2.
+ * @param[in]  seed Byte array with seed of length MLDSA_CRHBYTES.
+ */
+static void mld_sample_s1_s2(mld_polyvecl *s1, mld_polyveck *s2,
+                             const uint8_t seed[MLDSA_CRHBYTES])
+__contract__(
+  requires(memory_no_alias(s1, sizeof(mld_polyvecl)))
+  requires(memory_no_alias(s2, sizeof(mld_polyveck)))
+  requires(memory_no_alias(seed, MLDSA_CRHBYTES))
+  assigns(object_whole(s1), object_whole(s2))
+  ensures(forall(l0, 0, MLDSA_L, array_abs_bound(s1->vec[l0].coeffs, 0, MLDSA_N, MLDSA_ETA + 1)))
+  ensures(forall(k0, 0, MLDSA_K, array_abs_bound(s2->vec[k0].coeffs, 0, MLDSA_N, MLDSA_ETA + 1)))
+)
+{
+/* Sample short vectors s1 and s2 */
+#if defined(MLD_CONFIG_SERIAL_FIPS202_ONLY)
+  int i;
+  uint16_t nonce = 0;
+  /* Safety: The nonces are at most 14 (MLDSA_L + MLDSA_K - 1), and, hence, the
+   * casts are safe. */
+  for (i = 0; i < MLDSA_L; i++)
+  {
+    mld_poly_uniform_eta(&s1->vec[i], seed, (uint8_t)(nonce + i));
+  }
+  for (i = 0; i < MLDSA_K; i++)
+  {
+    mld_poly_uniform_eta(&s2->vec[i], seed, (uint8_t)(nonce + MLDSA_L + i));
+  }
+#else /* MLD_CONFIG_SERIAL_FIPS202_ONLY */
+#if MLD_CONFIG_PARAMETER_SET == 44
+  mld_poly_uniform_eta_4x(&s1->vec[0], &s1->vec[1], &s1->vec[2], &s1->vec[3],
+                          seed, 0, 1, 2, 3);
+  mld_poly_uniform_eta_4x(&s2->vec[0], &s2->vec[1], &s2->vec[2], &s2->vec[3],
+                          seed, 4, 5, 6, 7);
+#elif MLD_CONFIG_PARAMETER_SET == 65
+  mld_poly_uniform_eta_4x(&s1->vec[0], &s1->vec[1], &s1->vec[2], &s1->vec[3],
+                          seed, 0, 1, 2, 3);
+  mld_poly_uniform_eta_4x(&s1->vec[4], &s2->vec[0], &s2->vec[1],
+                          &s2->vec[2] /* irrelevant */, seed, 4, 5, 6,
+                          0xFF /* irrelevant */);
+  mld_poly_uniform_eta_4x(&s2->vec[2], &s2->vec[3], &s2->vec[4], &s2->vec[5],
+                          seed, 7, 8, 9, 10);
+#elif MLD_CONFIG_PARAMETER_SET == 87
+  mld_poly_uniform_eta_4x(&s1->vec[0], &s1->vec[1], &s1->vec[2], &s1->vec[3],
+                          seed, 0, 1, 2, 3);
+  mld_poly_uniform_eta_4x(&s1->vec[4], &s1->vec[5], &s1->vec[6],
+                          &s2->vec[0] /* irrelevant */, seed, 4, 5, 6,
+                          0xFF /* irrelevant */);
+  mld_poly_uniform_eta_4x(&s2->vec[0], &s2->vec[1], &s2->vec[2], &s2->vec[3],
+                          seed, 7, 8, 9, 10);
+  mld_poly_uniform_eta_4x(&s2->vec[4], &s2->vec[5], &s2->vec[6], &s2->vec[7],
+                          seed, 11, 12, 13, 14);
+#endif /* MLD_CONFIG_PARAMETER_SET == 87 */
+#endif /* !MLD_CONFIG_SERIAL_FIPS202_ONLY */
+}
+
+/**
+ * Compute t = A*s1hat + s2 row by row, decompose each row into t0[k] and
+ * t1[k] via power2round, and bit-pack t1[k] into pk_t1 and t0[k] into the
+ * t0_packed buffer. Used by both keygen and pk_from_sk.
+ *
+ * @spec{Partially implements @[FIPS204, Algorithm 22, pkEncode] (t1) and
+ * @[FIPS204, Algorithm 24, skEncode] (t0).}
+ *
+ * @param[out] pk_t1     Output buffer for packed t1 (size
+ *                       MLDSA_K * MLDSA_POLYT1_PACKEDBYTES; i.e. the t1
+ *                       region of pk).
+ * @param[out] t0_packed Output buffer for packed t0 (size
+ *                       MLDSA_K * MLDSA_POLYT0_PACKEDBYTES).
+ * @param[in]  s1hat     s1 in NTT domain.
+ * @param[in]  s2        s2.
+ * @param[in]  rho       Byte array containing seed rho.
+ * @param      context   Application context. Only present when
+ *                       MLD_CONFIG_CONTEXT_PARAMETER is defined; type set by
+ *                       MLD_CONFIG_CONTEXT_PARAMETER_TYPE.
+ *
+ * @return - 0: Success.
+ *         - MLD_ERR_OUT_OF_MEMORY: If MLD_CONFIG_CUSTOM_ALLOC_FREE is used and
+ *           an allocation via MLD_CUSTOM_ALLOC returned NULL.
+ */
+MLD_MUST_CHECK_RETURN_VALUE
+static int mld_compute_pack_t0_t1(
+    uint8_t pk_t1[MLDSA_K * MLDSA_POLYT1_PACKEDBYTES],
+    uint8_t t0_packed[MLDSA_K * MLDSA_POLYT0_PACKEDBYTES],
+    const mld_polyvecl *s1hat, const mld_polyveck *s2,
+    const uint8_t rho[MLDSA_SEEDBYTES],
+    MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+__contract__(
+  requires(memory_no_alias(pk_t1, MLDSA_K * MLDSA_POLYT1_PACKEDBYTES))
+  requires(memory_no_alias(t0_packed, MLDSA_K * MLDSA_POLYT0_PACKEDBYTES))
+  requires(memory_no_alias(s1hat, sizeof(mld_polyvecl)))
+  requires(memory_no_alias(s2, sizeof(mld_polyveck)))
+  requires(memory_no_alias(rho, MLDSA_SEEDBYTES))
+  requires(forall(l1, 0, MLDSA_L,
+    array_abs_bound(s1hat->vec[l1].coeffs, 0, MLDSA_N, MLD_NTT_BOUND)))
+  requires(forall(k2, 0, MLDSA_K,
+    array_bound(s2->vec[k2].coeffs, 0, MLDSA_N,
+      MLD_POLYETA_UNPACK_LOWER_BOUND, MLDSA_ETA + 1)))
+  assigns(memory_slice(pk_t1, MLDSA_K * MLDSA_POLYT1_PACKEDBYTES))
+  assigns(memory_slice(t0_packed, MLDSA_K * MLDSA_POLYT0_PACKEDBYTES))
+  ensures(return_value == 0 || return_value == MLD_ERR_OUT_OF_MEMORY))
+{
+  unsigned int k;
+  int ret;
+  MLD_ALLOC(mat, mld_polymat, 1, context);
+  MLD_ALLOC(t0k, mld_poly, 1, context);
+  MLD_ALLOC(t1k, mld_poly, 1, context);
+
+  if (mat == NULL || t0k == NULL || t1k == NULL)
+  {
+    ret = MLD_ERR_OUT_OF_MEMORY;
+    goto cleanup;
+  }
+
+  /* Expand matrix */
+  mld_polyvec_matrix_expand(mat, rho);
+
+  for (k = 0; k < MLDSA_K; k++)
+  __loop__(
+    assigns(k, memory_slice(pk_t1, MLDSA_K * MLDSA_POLYT1_PACKEDBYTES),
+            memory_slice(t0_packed, MLDSA_K * MLDSA_POLYT0_PACKEDBYTES),
+            memory_slice(t0k, sizeof(mld_poly)),
+            memory_slice(t1k, sizeof(mld_poly))
+            MLD_IF_REDUCE_RAM(, memory_slice(mat, sizeof(mld_polymat))))
+    invariant(k <= MLDSA_K)
+    decreases(MLDSA_K - k)
+  )
+  {
+    /* t0k = (A * s1hat)_k in NTT domain */
+    mld_polyvec_matrix_pointwise_montgomery_row(t0k, mat, s1hat, k);
+
+    /* t0k = invNTT(t0k) */
+    mld_poly_invntt_tomont(t0k);
+
+    /* t0k += s2[k] */
+    mld_poly_add(t0k, &s2->vec[k]);
+
+    /* Reference: The following reduction is not present in the reference
+     *            implementation. Omitting this reduction requires the output
+     *            of the invntt to be small enough such that the addition of
+     *            s2 does not result in absolute values >= MLDSA_Q. While our
+     *            C, x86_64, and AArch64 invntt implementations produce small
+     *            enough values for this to work out, it complicates the
+     *            bounds reasoning. We instead add an additional reduction,
+     *            and can consequently, relax the bounds requirements for the
+     *            invntt.
+     */
+    mld_poly_reduce(t0k);
+
+    /* Decompose into t1[k] and t0[k] (in place into t0k). */
+    mld_poly_caddq(t0k);
+    mld_poly_power2round(t1k, t0k, t0k);
+
+    /* Pack t1[k] into pk and t0[k] into the t0 output buffer. */
+    mld_polyt1_pack(pk_t1 + k * MLDSA_POLYT1_PACKEDBYTES, t1k);
+    mld_polyt0_pack(t0_packed + k * MLDSA_POLYT0_PACKEDBYTES, t0k);
+  }
+
+  ret = 0;
+cleanup:
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  MLD_FREE(t1k, mld_poly, 1, context);
+  MLD_FREE(t0k, mld_poly, 1, context);
+  MLD_FREE(mat, mld_polymat, 1, context);
+  return ret;
+}
+
+MLD_MUST_CHECK_RETURN_VALUE
+MLD_EXTERNAL_API
+int mld_sign_keypair_internal(uint8_t pk[MLDSA_CRYPTO_PUBLICKEYBYTES],
+                              uint8_t sk[MLDSA_CRYPTO_SECRETKEYBYTES],
+                              const uint8_t seed[MLDSA_SEEDBYTES],
+                              MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  int ret;
+  const uint8_t *rho, *rhoprime, *key;
+
+  MLD_ALLOC(seedbuf, uint8_t, 2 * MLDSA_SEEDBYTES + MLDSA_CRHBYTES, context);
+  MLD_ALLOC(inbuf, uint8_t, MLDSA_SEEDBYTES + 2, context);
+  MLD_ALLOC(tr, uint8_t, MLDSA_TRBYTES, context);
+  MLD_ALLOC(s1, mld_polyvecl, 1, context);
+  MLD_ALLOC(s2, mld_polyveck, 1, context);
+
+  if (seedbuf == NULL || inbuf == NULL || tr == NULL || s1 == NULL ||
+      s2 == NULL)
+  {
+    ret = MLD_ERR_OUT_OF_MEMORY;
+    goto cleanup;
+  }
+
+  /* Get randomness for rho, rhoprime and key */
+  mld_memcpy(inbuf, seed, MLDSA_SEEDBYTES);
+  inbuf[MLDSA_SEEDBYTES + 0] = MLDSA_K;
+  inbuf[MLDSA_SEEDBYTES + 1] = MLDSA_L;
+  mld_shake256(seedbuf, 2 * MLDSA_SEEDBYTES + MLDSA_CRHBYTES, inbuf,
+               MLDSA_SEEDBYTES + 2);
+  rho = seedbuf;
+  rhoprime = rho + MLDSA_SEEDBYTES;
+  key = rhoprime + MLDSA_CRHBYTES;
+
+  /* Constant time: rho is part of the public key and, hence, public. */
+  MLD_CT_TESTING_DECLASSIFY(rho, MLDSA_SEEDBYTES);
+
+  /* Sample s1 and s2 */
+  mld_sample_s1_s2(s1, s2, rhoprime);
+
+  /* Pack s1 into sk before NTT */
+  mld_pack_sk_s1(sk, s1);
+
+  /* NTT s1 in place to use as s1hat */
+  mld_polyvecl_ntt(s1);
+
+  /* Pack rho into pk */
+  mld_memcpy(pk + MLDSA_PK_RHO_OFFSET, rho, MLDSA_SEEDBYTES);
+
+  /* Compute t = A*s1hat + s2 row by row, decompose into t1/t0, and pack
+   * t1 into pk and t0 directly into the t0 region of sk. */
+  ret = mld_compute_pack_t0_t1(pk + MLDSA_PK_T1_OFFSET, sk + MLDSA_SK_T0_OFFSET,
+                               s1, s2, rho, context);
+  if (ret != 0)
+  {
+    goto cleanup;
+  }
+
+  /* Compute tr = H(pk) */
+  mld_shake256(tr, MLDSA_TRBYTES, pk, MLDSA_CRYPTO_PUBLICKEYBYTES);
+
+  /* Pack remaining secret key components (s1 and t0 already packed) */
+  mld_pack_sk_rho_key_tr_s2(sk, rho, tr, key, s2);
+
+  /* Constant time: pk is the public key, inherently public data */
+  MLD_CT_TESTING_DECLASSIFY(pk, MLDSA_CRYPTO_PUBLICKEYBYTES);
+
+cleanup:
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  MLD_FREE(s2, mld_polyveck, 1, context);
+  MLD_FREE(s1, mld_polyvecl, 1, context);
+  MLD_FREE(tr, uint8_t, MLDSA_TRBYTES, context);
+  MLD_FREE(inbuf, uint8_t, MLDSA_SEEDBYTES + 2, context);
+  MLD_FREE(seedbuf, uint8_t, 2 * MLDSA_SEEDBYTES + MLDSA_CRHBYTES, context);
+
+  /* Pairwise Consistency Test (PCT) @[FIPS140_3_IG, p.87] */
+  /* Do this after freeing all temporaries. */
+  if (ret == 0)
+  {
+    ret = mld_check_pct(pk, sk, context);
+  }
+
+  if (ret != 0)
+  {
+    /* Clear caller outputs on failure. */
+    mld_zeroize(pk, MLDSA_CRYPTO_PUBLICKEYBYTES);
+    mld_zeroize(sk, MLDSA_CRYPTO_SECRETKEYBYTES);
+  }
+
+  return ret;
+}
+
+#if !defined(MLD_CONFIG_CORE_API_ONLY)
+#if !defined(MLD_CONFIG_NO_RANDOMIZED_API)
+MLD_MUST_CHECK_RETURN_VALUE
+MLD_EXTERNAL_API
+int mld_sign_keypair(uint8_t pk[MLDSA_CRYPTO_PUBLICKEYBYTES],
+                     uint8_t sk[MLDSA_CRYPTO_SECRETKEYBYTES],
+                     MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  int ret;
+  MLD_ALLOC(seed, uint8_t, MLDSA_SEEDBYTES, context);
+
+  if (seed == NULL)
+  {
+    ret = MLD_ERR_OUT_OF_MEMORY;
+    goto cleanup;
+  }
+
+  if (mld_randombytes(seed, MLDSA_SEEDBYTES) != 0)
+  {
+    ret = MLD_ERR_RNG_FAIL;
+    goto cleanup;
+  }
+  MLD_CT_TESTING_SECRET(seed, MLDSA_SEEDBYTES);
+  ret = mld_sign_keypair_internal(pk, sk, seed, context);
+
+cleanup:
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  MLD_FREE(seed, uint8_t, MLDSA_SEEDBYTES, context);
+  return ret;
+}
+#endif /* !MLD_CONFIG_NO_RANDOMIZED_API */
+#endif /* !MLD_CONFIG_CORE_API_ONLY */
+#endif /* !MLD_CONFIG_NO_KEYPAIR_API */
+
+#if !defined(MLD_CONFIG_NO_SIGN_API) || !defined(MLD_CONFIG_NO_VERIFY_API)
+/**
+ * Abstracts application of SHAKE256 to one, two or three blocks of data,
+ * yielding a user-requested size of output.
+ *
+ * @param[out] out    Pointer to output.
+ * @param      outlen Requested output length in bytes.
+ * @param[in]  in1    Pointer to input block 1. Must NOT be NULL.
+ * @param      in1len Length of input in1 in bytes.
+ * @param[in]  in2    Pointer to input block 2. May be NULL if in2len == 0,
+ *                    in which case this block is ignored.
+ * @param      in2len Length of input in2 in bytes.
+ * @param[in]  in3    Pointer to input block 3. May be NULL if in3len == 0,
+ *                    in which case this block is ignored.
+ * @param      in3len Length of input in3 in bytes.
+ */
+static void mld_H(uint8_t *out, size_t outlen, const uint8_t *in1,
+                  size_t in1len, const uint8_t *in2, size_t in2len,
+                  const uint8_t *in3, size_t in3len)
+__contract__(
+  requires(in1len <= MLD_MAX_BUFFER_SIZE)
+  requires(in2len <= MLD_MAX_BUFFER_SIZE)
+  requires(in3len <= MLD_MAX_BUFFER_SIZE)
+  requires(outlen <= 8 * SHAKE256_RATE /* somewhat arbitrary bound */)
+  requires(memory_no_alias(in1, in1len))
+  requires(in2len == 0 || memory_no_alias(in2, in2len))
+  requires(in3len == 0 || memory_no_alias(in3, in3len))
+  requires(memory_no_alias(out, outlen))
+  assigns(memory_slice(out, outlen))
+)
+{
+  mld_shake256ctx state;
+  mld_shake256_init(&state);
+  mld_shake256_absorb(&state, in1, in1len);
+  if (in2len != 0)
+  {
+    mld_shake256_absorb(&state, in2, in2len);
+  }
+  if (in3len != 0)
+  {
+    mld_shake256_absorb(&state, in3, in3len);
+  }
+  mld_shake256_finalize(&state);
+  mld_shake256_squeeze(out, outlen, &state);
+  mld_shake256_release(&state);
+
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  mld_zeroize(&state, sizeof(state));
+}
+#endif /* !MLD_CONFIG_NO_SIGN_API || !MLD_CONFIG_NO_VERIFY_API */
+
+#if !defined(MLD_CONFIG_NO_SIGN_API)
+/* MLD_MAX_KAPPA (see params.h) bounds the rejection-sampling counter kappa;
+ * MLD_MAX_SIGNING_ATTEMPTS below turns that into a bound on attempts. */
+
+/**
+ * Compute z = y + s1*c, check that z has coefficients smaller than
+ * MLDSA_GAMMA1 - MLDSA_BETA, and pack z into the signature buffer.
+ *
+ * @reference{This function is inlined into mld_sign_signature in the
+ * reference implementation.}
+ *
+ * @param[in,out] sig   Output signature.
+ * @param[in]     cp    Challenge polynomial.
+ * @param[in]     s1hat Secret vector s1 in NTT domain.
+ * @param[in]     y     Masking vector y (or seed in REDUCE_RAM mode).
+ * @param[out]    z     Scratch polynomial for z computation.
+ * @param[out]    tmp   Scratch polynomial.
+ *
+ * @return - 0: Success (z has coefficients smaller than
+ *           MLDSA_GAMMA1 - MLDSA_BETA).
+ *         - MLD_ERR_FAIL: z rejected (norm check failed).
+ *         - MLD_ERR_OUT_OF_MEMORY: If MLD_CONFIG_CUSTOM_ALLOC_FREE is used and
+ *           an allocation via MLD_CUSTOM_ALLOC returned NULL.
+ */
+MLD_MUST_CHECK_RETURN_VALUE
+static int mld_compute_pack_z(uint8_t sig[MLDSA_CRYPTO_BYTES],
+                              const mld_poly *cp, const mld_sk_s1hat *s1hat,
+                              const mld_yvec *y, mld_poly *z, mld_poly *tmp)
+__contract__(
+  requires(memory_no_alias(sig, MLDSA_CRYPTO_BYTES))
+  requires(memory_no_alias(cp, sizeof(mld_poly)))
+  requires(memory_no_alias(s1hat, sizeof(mld_sk_s1hat)))
+  requires(memory_no_alias(y, sizeof(mld_yvec)))
+  requires(memory_no_alias(z, sizeof(mld_poly)))
+  requires(memory_no_alias(tmp, sizeof(mld_poly)))
+  requires(array_abs_bound(cp->coeffs, 0, MLDSA_N, MLD_NTT_BOUND))
+  MLD_IF_NOT_REDUCE_RAM(
+    requires(forall(k0, 0, MLDSA_L,
+      array_bound(y->vec.vec[k0].coeffs, 0, MLDSA_N, -(MLDSA_GAMMA1 - 1), MLDSA_GAMMA1 + 1)))
+    requires(forall(k1, 0, MLDSA_L, array_abs_bound(s1hat->vec.vec[k1].coeffs, 0, MLDSA_N, MLD_NTT_BOUND)))
+  )
+  MLD_IF_REDUCE_RAM(
+    requires(memory_no_alias(s1hat->packed, MLDSA_L * MLDSA_POLYETA_PACKEDBYTES))
+    requires(memory_no_alias(y->rhoprime, MLDSA_CRHBYTES))
+    requires(y->kappa <= MLD_MAX_KAPPA)
+  )
+  assigns(memory_slice(sig, MLDSA_CRYPTO_BYTES))
+  assigns(memory_slice(z, sizeof(mld_poly)))
+  assigns(memory_slice(tmp, sizeof(mld_poly)))
+  ensures(return_value == 0 || return_value == MLD_ERR_FAIL ||
+          return_value == MLD_ERR_OUT_OF_MEMORY)
+)
+{
+  unsigned int i;
+  uint32_t z_invalid;
+  for (i = 0; i < MLDSA_L; i++)
+  __loop__(
+    assigns(i, memory_slice(z, sizeof(mld_poly)),
+            memory_slice(tmp, sizeof(mld_poly)),
+            memory_slice(sig, MLDSA_CRYPTO_BYTES))
+    invariant(i <= MLDSA_L)
+    decreases(MLDSA_L - i)
+  )
+  {
+    mld_sk_s1hat_get_poly(z, s1hat, i);
+    mld_poly_pointwise_montgomery(z, cp);
+    mld_poly_invntt_tomont(z);
+    mld_yvec_get_poly(tmp, y, i);
+    mld_poly_add(z, tmp);
+    mld_poly_reduce(z);
+
+    z_invalid = mld_poly_chknorm(z, MLDSA_GAMMA1 - MLDSA_BETA);
+    /* Constant time: It is fine (and prohibitively expensive to avoid)
+     * to leak the result of the norm check and which polynomial in z caused a
+     * rejection. It would even be okay to leak which coefficient led to
+     * rejection as the candidate signature will be discarded anyway.
+     * See Section 5.5 of @[Round3_Spec]. */
+    MLD_CT_TESTING_DECLASSIFY(&z_invalid, sizeof(uint32_t));
+    if (z_invalid)
+    {
+      return MLD_ERR_FAIL; /* reject */
+    }
+    /* If z is valid, then its coefficients are bounded by
+     * MLDSA_GAMMA1 - MLDSA_BETA. This will be needed below
+     * to prove the pre-condition of pack_sig_z() */
+    mld_assert_abs_bound(z, MLDSA_N, (MLDSA_GAMMA1 - MLDSA_BETA));
+
+    /* After the norm check, the distribution of each coefficient of z is
+     * independent of the secret key and it can, hence, be considered
+     * public. It is, hence, okay to immediately pack it into the user-provided
+     * signature buffer. */
+    mld_pack_sig_z(sig, z, i);
+  }
+  return 0;
+}
+
+/* Effective bound on signing attempts: the configured bound
+ * MLD_CONFIG_MAX_SIGNING_ATTEMPTS (see mldsa_native_config.h) if set, otherwise
+ * the hard type-safety bound MLD_MAX_KAPPA / MLDSA_L (see MLD_MAX_KAPPA in
+ * params.h). */
+#if defined(MLD_CONFIG_MAX_SIGNING_ATTEMPTS)
+
+#if !defined(MLD_ALLOW_NONCOMPLIANT_SIGNING_BOUND) && \
+    MLD_CONFIG_MAX_SIGNING_ATTEMPTS < 821
+#error Bad configuration: MLD_CONFIG_MAX_SIGNING_ATTEMPTS must be >= 821 for FIPS 204 compliance @[FIPS204, Appendix C] @[FIPS204_UPDATES]
+#endif
+
+#if MLD_CONFIG_MAX_SIGNING_ATTEMPTS < 1
+#error Bad configuration: MLD_CONFIG_MAX_SIGNING_ATTEMPTS must be >= 1
+#endif
+
+#if MLD_CONFIG_MAX_SIGNING_ATTEMPTS > MLD_MAX_KAPPA / MLDSA_L
+#error Bad configuration: MLD_CONFIG_MAX_SIGNING_ATTEMPTS exceeds the maximum allowed value.
+#endif
+
+#define MLD_MAX_SIGNING_ATTEMPTS MLD_CONFIG_MAX_SIGNING_ATTEMPTS
+#else /* MLD_CONFIG_MAX_SIGNING_ATTEMPTS */
+#define MLD_MAX_SIGNING_ATTEMPTS (MLD_MAX_KAPPA / MLDSA_L)
+#endif /* !MLD_CONFIG_MAX_SIGNING_ATTEMPTS */
+
+MLD_MUST_CHECK_RETURN_VALUE
+static MLD_INLINE uint16_t mld_get_max_signing_attempts(void)
+__contract__(
+  ensures(return_value >= 1)
+  ensures(return_value <= MLD_MAX_KAPPA / MLDSA_L)
+)
+{
+  /* cassert(0) ensures CBMC uses the contract rather than inlining the body,
+   * keeping proofs agnostic of the configured value. */
+  cassert(0);
+  return MLD_MAX_SIGNING_ATTEMPTS;
+}
+
+/**
+ * Attempt to generate a single signature: one iteration of the
+ * ML-DSA.Sign_internal rejection-sampling loop.
+ *
+ * @spec{Implements one iteration of the rejection-sampling loop body of
+ * @[FIPS204, Algorithm 7, ML-DSA.Sign_internal] (lines 11-30) plus, on success,
+ * the sigEncode step (line 33). The per-signature setup (Algorithm 7 lines 1-7:
+ * skDecode, NTT of s1/s2/t0, ExpandA, and computation of mu and rhoprime) and
+ * the loop itself (lines 8-10, 31-32) live in the caller
+ * mld_sign_signature_internal; kappa is this iteration's counter, used to
+ * sample y.}
+ *
+ * @reference{This code differs from the reference implementation in that it
+ * factors out the core signature generation step into a distinct function
+ * here in order to improve efficiency of CBMC proof.}
+ *
+ * @param[out] sig      Pointer to output signature.
+ * @param[in]  mu       Pointer to message or hash of exactly MLDSA_CRHBYTES
+ *                      bytes.
+ * @param[in]  rhoprime Pointer to randomness seed.
+ * @param      kappa    Counter for this iteration (= attempt*MLDSA_L).
+ * @param[in]  mat      Expanded matrix.
+ * @param[in]  s1hat    Secret vector s1 in NTT domain.
+ * @param[in]  s2hat    Secret vector s2 in NTT domain.
+ * @param[in]  t0hat    Vector t0 in NTT domain.
+ * @param      context  Application context. Only present when
+ *                      MLD_CONFIG_CONTEXT_PARAMETER is defined; type set by
+ *                      MLD_CONFIG_CONTEXT_PARAMETER_TYPE.
+ *
+ * @return - 0: Signature generation succeeded.
+ *         - MLD_ERR_FAIL: Signature rejected (norm check failed).
+ *         - MLD_ERR_OUT_OF_MEMORY: If MLD_CONFIG_CUSTOM_ALLOC_FREE is used and
+ *           an allocation via MLD_CUSTOM_ALLOC returned NULL.
+ */
+MLD_MUST_CHECK_RETURN_VALUE
+/* NOLINTNEXTLINE(readability-function-cognitive-complexity) */
+static int mld_attempt_signature_generation(
+    uint8_t sig[MLDSA_CRYPTO_BYTES], const uint8_t *mu,
+    const uint8_t rhoprime[MLDSA_CRHBYTES], uint16_t kappa, mld_polymat *mat,
+    const mld_sk_s1hat *s1hat, const mld_sk_s2hat *s2hat,
+    const mld_sk_t0hat *t0hat, MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+__contract__(
+  requires(memory_no_alias(sig, MLDSA_CRYPTO_BYTES))
+  requires(memory_no_alias(mu, MLDSA_CRHBYTES))
+  requires(memory_no_alias(rhoprime, MLDSA_CRHBYTES))
+  requires(memory_no_alias(mat, sizeof(mld_polymat)))
+  requires(memory_no_alias(s1hat, sizeof(mld_sk_s1hat)))
+  requires(memory_no_alias(s2hat, sizeof(mld_sk_s2hat)))
+  requires(memory_no_alias(t0hat, sizeof(mld_sk_t0hat)))
+  requires(kappa <= MLD_MAX_KAPPA)
+  MLD_IF_NOT_REDUCE_RAM(
+    requires(forall(k1, 0, MLDSA_K, forall(l1, 0, MLDSA_L,
+                                           array_bound(mat->vec[k1].vec[l1].coeffs, 0, MLDSA_N, 0, MLDSA_Q))))
+    requires(forall(k2, 0, MLDSA_K, array_abs_bound(t0hat->vec.vec[k2].coeffs, 0, MLDSA_N, MLD_NTT_BOUND)))
+    requires(forall(k3, 0, MLDSA_L, array_abs_bound(s1hat->vec.vec[k3].coeffs, 0, MLDSA_N, MLD_NTT_BOUND)))
+    requires(forall(k4, 0, MLDSA_K, array_abs_bound(s2hat->vec.vec[k4].coeffs, 0, MLDSA_N, MLD_NTT_BOUND)))
+  )
+  MLD_IF_REDUCE_RAM(
+    requires(memory_no_alias(s1hat->packed, MLDSA_L * MLDSA_POLYETA_PACKEDBYTES))
+    requires(memory_no_alias(s2hat->packed, MLDSA_K * MLDSA_POLYETA_PACKEDBYTES))
+    requires(memory_no_alias(t0hat->packed, MLDSA_K * MLDSA_POLYT0_PACKEDBYTES))
+  )
+  assigns(memory_slice(sig, MLDSA_CRYPTO_BYTES))
+  MLD_IF_REDUCE_RAM(
+    assigns(memory_slice(mat, sizeof(mld_polymat)))
+  )
+  ensures(return_value == 0 || return_value == MLD_ERR_FAIL ||
+          return_value == MLD_ERR_OUT_OF_MEMORY)
+)
+{
+  unsigned int k;
+  uint32_t w0_invalid, h_invalid;
+  int ret;
+
+  typedef union
+  {
+    mld_polyveck w1;
+    mld_polyvecl tmp;
+  } w1tmp_u;
+  mld_polyveck *w1;
+  mld_polyvecl *tmp;
+
+  MLD_ALLOC(challenge_bytes, uint8_t, MLDSA_CTILDEBYTES, context);
+  MLD_ALLOC(y, mld_yvec, 1, context);
+  MLD_ALLOC(z, mld_poly, 1, context);
+  MLD_ALLOC(w1tmp, w1tmp_u, 1, context);
+  MLD_ALLOC(w0, mld_polyveck, 1, context);
+  MLD_ALLOC(cp, mld_poly, 1, context);
+  MLD_ALLOC(t, mld_poly, 1, context);
+
+  if (challenge_bytes == NULL || y == NULL || z == NULL || w1tmp == NULL ||
+      w0 == NULL || cp == NULL || t == NULL)
+  {
+    ret = MLD_ERR_OUT_OF_MEMORY;
+    goto cleanup;
+  }
+  w1 = &w1tmp->w1;
+  tmp = &w1tmp->tmp;
+
+  /* @[FIPS204, Algorithm 7, line 11] y <- ExpandMask(rhoprime, kappa). */
+  mld_yvec_init(y, rhoprime, kappa);
+
+  /* @[FIPS204, Algorithm 7, line 12] w <- invNTT(A_hat o NTT(y)). This call
+   * performs the whole line: it NTTs y, accumulates the pointwise product with
+   * A_hat, and applies the inverse NTT. In REDUCE_RAM mode the y sampling is
+   * fused into the same pass. */
+  mld_polyvec_matrix_pointwise_montgomery_yvec(w0, mat, y, tmp);
+
+  /* @[FIPS204, Algorithm 7, line 13] w1 <- HighBits(w), here together with the
+   * low part: Decompose yields w = 2*GAMMA2*w1 + w0, keeping both w1 and w0
+   * (w0 is reused below in the line-21/26 alternative, see further down). */
+  mld_polyveck_caddq(w0);
+  mld_polyveck_decompose(w1, w0);
+
+  /* @[FIPS204, Algorithm 7, line 15] ctilde <- H(mu || w1Encode(w1), lambda/4).
+   * w1Encode(w1) is packed into the w1 region of sig (mld_polyveck_pack_w1),
+   * then absorbed by H together with mu. */
+  mld_polyveck_pack_w1(sig, w1);
+
+  mld_H(challenge_bytes, MLDSA_CTILDEBYTES, mu, MLDSA_CRHBYTES, sig,
+        MLDSA_K * MLDSA_POLYW1_PACKEDBYTES, NULL, 0);
+  /* Constant time: Leaking challenge_bytes does not reveal any information
+   * about the secret key as H() is modelled as random oracle.
+   * This also applies to challenges for rejected signatures.
+   * See Section 5.5 of @[Round3_Spec]. */
+  MLD_CT_TESTING_DECLASSIFY(challenge_bytes, MLDSA_CTILDEBYTES);
+  /* @[FIPS204, Algorithm 7, line 16] c <- SampleInBall(ctilde) and
+   * @[FIPS204, Algorithm 7, line 17] c_hat <- NTT(c). */
+  mld_poly_challenge(cp, challenge_bytes);
+  mld_poly_ntt(cp);
+
+  /* @[FIPS204, Algorithm 7, lines 18+20] cs1 <- invNTT(c_hat o s1_hat) and
+   * z <- y + cs1, followed by the line-23 norm check ||z||_inf >= GAMMA1 -
+   * BETA. mld_compute_pack_z fuses all three per polynomial and, on success,
+   * packs z into sig; it returns MLD_ERR_FAIL if the norm check rejects z. */
+  ret = mld_compute_pack_z(sig, cp, s1hat, y, t, z);
+  if (ret != 0)
+  {
+    goto cleanup;
+  }
+
+  /* The remaining steps realize @[FIPS204, Algorithm 7, lines 21-28] (the
+   * low-bits norm check and the hint h) via the faster alternative formulation
+   * of @[Round3_Spec, Section 5.1]. @[FIPS204] explicitly permits this: the
+   * note accompanying Algorithm 7 states that the validity checks on z and the
+   * computation of h may instead be implemented "as described in Section 5.1 of
+   * [6]", and that reference is @[Round3_Spec, Section 5.1].
+   *
+   * The loop below builds w0 - cs2 + ct0 in place in w0; w1 is unmodified, and
+   * is HighBits(w) from line 13. Those are the inputs to the streamlined
+   * computation of MakeHint explained below.
+   *
+   * Low-bits norm check:
+   *   @[FIPS204, Algorithm 7, line 21] computes r0 = LowBits(w - cs2) and line
+   *   23 rejects when ||r0||_inf >= GAMMA2 - BETA. By @[Round3_Spec, Section
+   *   5.1] (Lemma 3), this line-23 check on r0 = LowBits(w - cs2) is implied by
+   *   ||w0 - cs2||_inf < GAMMA2 - BETA, where w0 is the low part of w. In our
+   *   context, w0 already holds the low part of w from the line-13 Decompose;
+   *   after subtracting cs2 from it in place, the mld_poly_chknorm(w0, GAMMA2 -
+   *   BETA) call below is exactly that check.
+   *
+   * Hint:
+   *   @[FIPS204, Algorithm 7, line 26] sets h = MakeHint(-ct0, w - cs2 + ct0),
+   *   and line 28 rejects when ||ct0||_inf >= GAMMA2 or h has more than OMEGA
+   *   nonzero coefficients. @[Round3_Spec, Section 5.1] provides the following
+   *   alternative description for MakeHint(-ct0, w - cs2 + ct0): a hint bit is
+   *   zero exactly when the coefficient of w0 - cs2 + ct0 lies in
+   *   (-GAMMA2, GAMMA2], or equals -GAMMA2 while the matching w1 coefficient is
+   *   zero (the Decompose border case), and is set otherwise. This equivalence
+   *   is precisely what mld_pack_sig_h -> mld_make_hint compute from w0
+   *   (= w0 - cs2 + ct0) and w1. The line-28 ||ct0||_inf >= GAMMA2 check is the
+   *   mld_poly_chknorm(z, GAMMA2) call on ct0 below; the weight bound is
+   *   enforced by mld_pack_sig_h.
+   *
+   * Building w0 per-component and checking norms incrementally also avoids
+   * allocating a full polyveck for h. */
+  for (k = 0; k < MLDSA_K; k++)
+  __loop__(
+    assigns(k,
+            object_whole(z),
+            object_whole(w0))
+    invariant(k <= MLDSA_K)
+    invariant(forall(k0, k, MLDSA_K,
+      array_abs_bound(w0->vec[k0].coeffs, 0, MLDSA_N, MLDSA_GAMMA2 + 1)))
+    decreases(MLDSA_K - k)
+  )
+  {
+    /* @[FIPS204, Algorithm 7, line 19] cs2[k] <- invNTT(c_hat o s2_hat)[k],
+     * then subtract from w0[k] to form (w0 - cs2)[k]. */
+    mld_sk_s2hat_get_poly(z, s2hat, k);
+    mld_poly_pointwise_montgomery(z, cp);
+    mld_poly_invntt_tomont(z);
+
+    mld_poly_sub(&w0->vec[k], z);
+    mld_poly_reduce(&w0->vec[k]);
+
+    /* Low-bits norm check (see block comment above): the line-23 check on
+     * r0 = LowBits(w - cs2) holds via ||w0 - cs2||_inf < GAMMA2 - BETA. */
+    w0_invalid = mld_poly_chknorm(&w0->vec[k], MLDSA_GAMMA2 - MLDSA_BETA);
+    /* Constant time: w0_invalid may be leaked - see comment for z_invalid. */
+    MLD_CT_TESTING_DECLASSIFY(&w0_invalid, sizeof(uint32_t));
+    if (w0_invalid)
+    {
+      ret = MLD_ERR_FAIL; /* reject */
+      goto cleanup;
+    }
+
+    /* @[FIPS204, Algorithm 7, line 25] ct0[k] <- invNTT(c_hat o t0_hat)[k]. */
+    mld_sk_t0hat_get_poly(z, t0hat, k);
+    mld_poly_pointwise_montgomery(z, cp);
+    mld_poly_invntt_tomont(z);
+    mld_poly_reduce(z);
+
+    /* @[FIPS204, Algorithm 7, line 28] reject when ||ct0||_inf >= GAMMA2 (the
+     * second part, the OMEGA weight bound, is enforced by mld_pack_sig_h). */
+    h_invalid = mld_poly_chknorm(z, MLDSA_GAMMA2);
+    /* Constant time: h_invalid may be leaked - see comment for z_invalid. */
+    MLD_CT_TESTING_DECLASSIFY(&h_invalid, sizeof(uint32_t));
+    if (h_invalid)
+    {
+      ret = MLD_ERR_FAIL; /* reject */
+      goto cleanup;
+    }
+
+    /* Add ct0[k] to (w0 - cs2)[k], leaving (w0 - cs2 + ct0)[k] in w0[k] -- the
+     * MakeHint input prepared for mld_pack_sig_h (see block comment above). */
+    mld_poly_add(&w0->vec[k], z);
+  }
+
+  /* Constant time: At this point all norm checks have passed and we, hence,
+   * know that the signature does not leak any secret information.
+   * Consequently, any value that can be computed from the signature and public
+   * key is considered public.
+   * w0 and w1 are public as they can be computed from Az - ct = \alpha w1 + w0.
+   * h=c*t0 is public as both c and t0 are considered public.
+   * While t0 is not part of the public key, it can be reconstructed from
+   * a small number of signatures and need not be regarded as secret
+   * (see @[FIPS204, Section 6.1]).
+   */
+  MLD_CT_TESTING_DECLASSIFY(w0, sizeof(*w0));
+  MLD_CT_TESTING_DECLASSIFY(w1, sizeof(*w1));
+
+  /* @[FIPS204, Algorithm 7, line 33] sigEncode(ctilde, z mod+/- q, h) is split
+   * across three calls: z was already packed by mld_compute_pack_z, this call
+   * packs ctilde, and mld_pack_sig_h below packs the hint h. */
+  mld_pack_sig_c(sig, challenge_bytes);
+
+  /* @[FIPS204, Algorithm 7, line 26] h <- MakeHint(-ct0, w - cs2 + ct0),
+   * computed from (w0 = w0 - cs2 + ct0, w1) as described in the block comment
+   * above, and packed as the h component of the line-33 sigEncode. Returns
+   * MLD_ERR_FAIL if h would exceed OMEGA nonzero coefficients (the remaining
+   * part of the line-28 check), in which case we reject. */
+  ret = mld_pack_sig_h(sig, w0, w1);
+  if (ret != 0)
+  {
+    goto cleanup;
+  }
+
+  /* Constant time: At this point it is clear that the signature is valid - it
+   * can, hence, be considered public. */
+  MLD_CT_TESTING_DECLASSIFY(sig, MLDSA_CRYPTO_BYTES);
+  ret = 0; /* success */
+
+cleanup:
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  MLD_FREE(t, mld_poly, 1, context);
+  MLD_FREE(cp, mld_poly, 1, context);
+  MLD_FREE(w0, mld_polyveck, 1, context);
+  MLD_FREE(w1tmp, w1tmp_u, 1, context);
+  MLD_FREE(z, mld_poly, 1, context);
+  MLD_FREE(y, mld_yvec, 1, context);
+  MLD_FREE(challenge_bytes, uint8_t, MLDSA_CTILDEBYTES, context);
+
+  return ret;
+}
+MLD_MUST_CHECK_RETURN_VALUE
+MLD_EXTERNAL_API
+int mld_sign_signature_internal(uint8_t sig[MLDSA_CRYPTO_BYTES],
+                                const uint8_t *m, size_t mlen,
+                                const uint8_t *pre, size_t prelen,
+                                const uint8_t rnd[MLDSA_RNDBYTES],
+                                const uint8_t sk[MLDSA_CRYPTO_SECRETKEYBYTES],
+                                int externalmu,
+                                MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  int ret;
+  uint8_t *rho, *tr, *key, *mu, *rhoprime;
+  uint16_t attempt;
+  const uint16_t max_signing_attempts = mld_get_max_signing_attempts();
+  MLD_ALLOC(seedbuf, uint8_t,
+            2 * MLDSA_SEEDBYTES + MLDSA_TRBYTES + 2 * MLDSA_CRHBYTES, context);
+  MLD_ALLOC(mat, mld_polymat, 1, context);
+  MLD_ALLOC(s1hat, mld_sk_s1hat, 1, context);
+  MLD_ALLOC(t0hat, mld_sk_t0hat, 1, context);
+  MLD_ALLOC(s2hat, mld_sk_s2hat, 1, context);
+
+  if (seedbuf == NULL || mat == NULL || s1hat == NULL || t0hat == NULL ||
+      s2hat == NULL)
+  {
+    ret = MLD_ERR_OUT_OF_MEMORY;
+    goto cleanup;
+  }
+
+  /* If a resume hook is configured via MLD_CONFIG_SIGN_HOOK_RESUME, it provides
+   * the attempt to resume from after an earlier pause. Otherwise, we start at
+   * 0. Clamp to max_signing_attempts. */
+  attempt = mld_sign_resume(context);
+  if (attempt > max_signing_attempts)
+  {
+    attempt = max_signing_attempts;
+  }
+
+  rho = seedbuf;
+  tr = rho + MLDSA_SEEDBYTES;
+  key = tr + MLDSA_TRBYTES;
+  mu = key + MLDSA_SEEDBYTES;
+  rhoprime = mu + MLDSA_CRHBYTES;
+  /* @[FIPS204, Algorithm 7, line 1] (rho, K, tr, s1, s2, t0) <- skDecode(sk)
+   * and @[FIPS204, Algorithm 7, lines 2-4] s1_hat/s2_hat/t0_hat <- NTT(...):
+   * mld_unpack_sk returns s1hat, s2hat, t0hat already in NTT domain. The spec's
+   * private random seed K is held in the local variable key. */
+  mld_unpack_sk(rho, tr, key, t0hat, s1hat, s2hat, sk);
+
+  if (!externalmu)
+  {
+    /* @[FIPS204, Algorithm 7, line 6] mu <- H(BytesToBits(tr) || M', 64). */
+    mld_H(mu, MLDSA_CRHBYTES, tr, MLDSA_TRBYTES, pre, prelen, m, mlen);
+  }
+  else
+  {
+    /* mu has been provided directly (external-mu variant; line 6 done by the
+     * caller in a separate cryptographic module). */
+    mld_memcpy(mu, m, MLDSA_CRHBYTES);
+  }
+
+  /* @[FIPS204, Algorithm 7, line 7] rhoprime <- H(K || rnd || mu, 64). */
+  mld_H(rhoprime, MLDSA_CRHBYTES, key, MLDSA_SEEDBYTES, rnd, MLDSA_RNDBYTES, mu,
+        MLDSA_CRHBYTES);
+
+  /* Constant time: rho is part of the public key and, hence, public. */
+  MLD_CT_TESTING_DECLASSIFY(rho, MLDSA_SEEDBYTES);
+  /* @[FIPS204, Algorithm 7, line 5] A_hat <- ExpandA(rho). */
+  mld_polyvec_matrix_expand(mat, rho);
+
+  /* @[FIPS204, Algorithm 7, lines 8-10 and 31-32] the rejection-sampling loop,
+   * tracked by attempt (kappa = attempt*MLDSA_L). Each iteration's body (lines
+   * 11-30) plus, on success, the line-33 sigEncode are performed by
+   * mld_attempt_signature_generation. */
+
+  /* Reference: the reference loops unboundedly; we instead iterate over the
+   * bounded range [0, max_signing_attempts) for predictable termination.
+   * A success or fatal error exits via goto cleanup; running to completion
+   * means every attempt was rejected; with a FIPS compliant choice of
+   * MLD_CONFIG_MAX_SIGNING_ATTEMPTS, this should never happen. */
+  for (; attempt < max_signing_attempts; attempt++)
+  __loop__(
+    MLD_IF_NOT_REDUCE_RAM(
+      assigns(attempt, ret, memory_slice(sig, MLDSA_CRYPTO_BYTES))
+    )
+    MLD_IF_REDUCE_RAM(
+      assigns(attempt, ret, memory_slice(sig, MLDSA_CRYPTO_BYTES),
+              memory_slice(mat, sizeof(mld_polymat)))
+    )
+    invariant(attempt <= max_signing_attempts)
+
+    /* t0, s1, s2, and mat are initialized above and are NOT changed by this */
+    /* loop. We can therefore re-assert their bounds here as part of the     */
+    /* loop invariant. This makes proof noticeably faster with CBMC          */
+    MLD_IF_NOT_REDUCE_RAM(
+      invariant(forall(k1, 0, MLDSA_K, forall(l1, 0, MLDSA_L,
+                array_bound(mat->vec[k1].vec[l1].coeffs, 0, MLDSA_N, 0, MLDSA_Q))))
+      invariant(forall(k2, 0, MLDSA_K, array_abs_bound(t0hat->vec.vec[k2].coeffs, 0, MLDSA_N, MLD_NTT_BOUND)))
+      invariant(forall(k3, 0, MLDSA_L, array_abs_bound(s1hat->vec.vec[k3].coeffs, 0, MLDSA_N, MLD_NTT_BOUND)))
+      invariant(forall(k4, 0, MLDSA_K, array_abs_bound(s2hat->vec.vec[k4].coeffs, 0, MLDSA_N, MLD_NTT_BOUND)))
+    )
+    decreases(max_signing_attempts - attempt)
+  )
+  {
+    /* Safety: attempt < max_signing_attempts <= MLD_MAX_KAPPA / MLDSA_L, so
+     * kappa <= MLD_MAX_KAPPA and the cast is safe. */
+    const uint16_t kappa = (uint16_t)(attempt * MLDSA_L);
+
+    /* Query configurable signing hook whether signing should be paused.
+     * This is skipped by default and only used if the user sets the
+     * configuration option MLD_CONFIG_SIGN_HOOK_ATTEMPT. */
+    if (mld_sign_attempt(attempt, context) != 0)
+    {
+      ret = MLD_ERR_SIGNING_PAUSED;
+      goto cleanup;
+    }
+
+    ret = mld_attempt_signature_generation(sig, mu, rhoprime, kappa, mat, s1hat,
+                                           s2hat, t0hat, context);
+
+    /* Decide whether to keep trying based on the return value:
+     *  - ret == 0: a valid signature was produced; we are done.
+     *  - ret == MLD_ERR_FAIL: this candidate was rejected by one of the norm
+     *    or hint checks. We continue the loop and try again with the next
+     *    nonce.
+     *  - any other value (e.g. MLD_ERR_OUT_OF_MEMORY): an unrecoverable error
+     *    occurred, so we propagate it to the caller. */
+    if (ret == 0)
+    {
+      /* Signing succeeded: record the attempt that succeeded. No-op in the
+       * default build. */
+      mld_sign_finish(attempt, context);
+      goto cleanup;
+    }
+    if (ret != MLD_ERR_FAIL)
+    {
+      goto cleanup;
+    }
+  }
+
+  /* Loop ran to completion: all attempts rejected, budget exhausted.
+   * This should never happen with a FIPS compliant choice of
+   * MLD_CONFIG_MAX_SIGNING_ATTEMPTS. */
+  ret = MLD_ERR_SIGN_ATTEMPTS_EXHAUSTED;
+
+cleanup:
+
+  if (ret != 0)
+  {
+    /* To be on the safe-side, we zeroize the signature buffer. */
+    mld_zeroize(sig, MLDSA_CRYPTO_BYTES);
+  }
+
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  MLD_FREE(s2hat, mld_sk_s2hat, 1, context);
+  MLD_FREE(t0hat, mld_sk_t0hat, 1, context);
+  MLD_FREE(s1hat, mld_sk_s1hat, 1, context);
+  MLD_FREE(mat, mld_polymat, 1, context);
+  MLD_FREE(seedbuf, uint8_t,
+           2 * MLDSA_SEEDBYTES + MLDSA_TRBYTES + 2 * MLDSA_CRHBYTES, context);
+  return ret;
+}
+
+#if !defined(MLD_CONFIG_CORE_API_ONLY)
+#if !defined(MLD_CONFIG_NO_RANDOMIZED_API)
+MLD_MUST_CHECK_RETURN_VALUE
+MLD_EXTERNAL_API
+int mld_sign_signature(uint8_t sig[MLDSA_CRYPTO_BYTES], const uint8_t *m,
+                       size_t mlen, const uint8_t *ctx, size_t ctxlen,
+                       const uint8_t sk[MLDSA_CRYPTO_SECRETKEYBYTES],
+                       MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  size_t pre_len;
+  int ret;
+  MLD_ALLOC(pre, uint8_t, MLD_DOMAIN_SEPARATION_MAX_BYTES, context);
+  MLD_ALLOC(rnd, uint8_t, MLDSA_RNDBYTES, context);
+
+  if (pre == NULL || rnd == NULL)
+  {
+    ret = MLD_ERR_OUT_OF_MEMORY;
+    goto cleanup;
+  }
+
+  /* Prepare domain separation prefix for pure ML-DSA */
+  pre_len = mld_prepare_domain_separation_prefix(pre, NULL, 0, ctx, ctxlen,
+                                                 MLD_PREHASH_NONE);
+  if (pre_len == 0)
+  {
+    ret = MLD_ERR_INVALID_ARG;
+    goto cleanup;
+  }
+
+  /* Randomized variant of ML-DSA. If you need the deterministic variant,
+   * call mld_sign_signature_internal directly with all-zero rnd. */
+  if (mld_randombytes(rnd, MLDSA_RNDBYTES) != 0)
+  {
+    ret = MLD_ERR_RNG_FAIL;
+    goto cleanup;
+  }
+  MLD_CT_TESTING_SECRET(rnd, MLDSA_RNDBYTES);
+
+  ret = mld_sign_signature_internal(sig, m, mlen, pre, pre_len, rnd, sk, 0,
+                                    context);
+
+cleanup:
+  if (ret != 0)
+  {
+    /* To be on the safe-side, make sure sig has a well-defined value, even in
+     * the case of error.
+     *
+     * If we come from mld_sign_signature_internal, this is redundant, but the
+     * error case should not be the norm, and the added cost of the zeroization
+     * insignificant. */
+    mld_zeroize(sig, MLDSA_CRYPTO_BYTES);
+  }
+
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  MLD_FREE(rnd, uint8_t, MLDSA_RNDBYTES, context);
+  MLD_FREE(pre, uint8_t, MLD_DOMAIN_SEPARATION_MAX_BYTES, context);
+
+  return ret;
+}
+#endif /* !MLD_CONFIG_NO_RANDOMIZED_API */
+
+#if !defined(MLD_CONFIG_NO_RANDOMIZED_API)
+MLD_MUST_CHECK_RETURN_VALUE
+MLD_EXTERNAL_API
+int mld_sign_signature_extmu(uint8_t sig[MLDSA_CRYPTO_BYTES],
+                             const uint8_t mu[MLDSA_CRHBYTES],
+                             const uint8_t sk[MLDSA_CRYPTO_SECRETKEYBYTES],
+                             MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  int ret;
+  MLD_ALLOC(rnd, uint8_t, MLDSA_RNDBYTES, context);
+
+  if (rnd == NULL)
+  {
+    ret = MLD_ERR_OUT_OF_MEMORY;
+    goto cleanup;
+  }
+
+  /* Randomized variant of ML-DSA. If you need the deterministic variant,
+   * call mld_sign_signature_internal directly with all-zero rnd. */
+  if (mld_randombytes(rnd, MLDSA_RNDBYTES) != 0)
+  {
+    ret = MLD_ERR_RNG_FAIL;
+    goto cleanup;
+  }
+  MLD_CT_TESTING_SECRET(rnd, MLDSA_RNDBYTES);
+
+  ret = mld_sign_signature_internal(sig, mu, MLDSA_CRHBYTES, NULL, 0, rnd, sk,
+                                    1, context);
+
+cleanup:
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  MLD_FREE(rnd, uint8_t, MLDSA_RNDBYTES, context);
+
+  return ret;
+}
+#endif /* !MLD_CONFIG_NO_RANDOMIZED_API */
+#endif /* !MLD_CONFIG_CORE_API_ONLY */
+#endif /* !MLD_CONFIG_NO_SIGN_API */
+
+#if !defined(MLD_CONFIG_NO_VERIFY_API)
+MLD_MUST_CHECK_RETURN_VALUE
+MLD_EXTERNAL_API
+int mld_sign_verify_internal(const uint8_t sig[MLDSA_CRYPTO_BYTES],
+                             const uint8_t *m, size_t mlen, const uint8_t *pre,
+                             size_t prelen,
+                             const uint8_t pk[MLDSA_CRYPTO_PUBLICKEYBYTES],
+                             int externalmu,
+                             MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  int ret, cmp;
+  unsigned int i;
+
+  MLD_ALLOC(buf, uint8_t, (MLDSA_K * MLDSA_POLYW1_PACKEDBYTES), context);
+  MLD_ALLOC(mu, uint8_t, MLDSA_CRHBYTES, context);
+  MLD_ALLOC(c, uint8_t, MLDSA_CTILDEBYTES, context);
+  MLD_ALLOC(c2, uint8_t, MLDSA_CTILDEBYTES, context);
+  MLD_ALLOC(z, mld_polyvecl, 1, context);
+  MLD_ALLOC(cp, mld_poly, 1, context);
+  MLD_ALLOC(mat, mld_polymat, 1, context);
+  MLD_ALLOC(w1, mld_poly, 1, context);
+  MLD_ALLOC(tmp, mld_poly, 1, context);
+
+  if (buf == NULL || mu == NULL || c == NULL || c2 == NULL || z == NULL ||
+      cp == NULL || mat == NULL || w1 == NULL || tmp == NULL)
+  {
+    ret = MLD_ERR_OUT_OF_MEMORY;
+    goto cleanup;
+  }
+
+  mld_memcpy(c, sig, MLDSA_CTILDEBYTES);
+  mld_polyvecl_unpack_z(z, sig + MLDSA_SIG_Z_OFFSET);
+
+  /* mld_polyvecl_chknorm signals failure through a single non-zero error code
+   * that's not yet aligned with MLD_ERR_XXX. A norm-check failure here means
+   * the signature is invalid, so map it to MLD_ERR_INVALID_SIGNATURE. */
+  if (mld_polyvecl_chknorm(z, MLDSA_GAMMA1 - MLDSA_BETA))
+  {
+    ret = MLD_ERR_INVALID_SIGNATURE;
+    goto cleanup;
+  }
+
+  if (!externalmu)
+  {
+    /* Compute CRH(H(rho, t1), pre, msg) */
+    MLD_ALIGN uint8_t hpk[MLDSA_CRHBYTES];
+    mld_H(hpk, MLDSA_TRBYTES, pk, MLDSA_CRYPTO_PUBLICKEYBYTES, NULL, 0, NULL,
+          0);
+    mld_H(mu, MLDSA_CRHBYTES, hpk, MLDSA_TRBYTES, pre, prelen, m, mlen);
+
+    /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+    mld_zeroize(hpk, sizeof(hpk));
+  }
+  else
+  {
+    /* mu has been provided directly */
+    mld_memcpy(mu, m, MLDSA_CRHBYTES);
+  }
+
+  /* Matrix-vector multiplication and per-row reconstruction of w1. */
+  mld_polyvecl_ntt(z);
+  mld_polyvec_matrix_expand(mat, pk);
+  mld_poly_challenge(cp, c);
+  mld_poly_ntt(cp);
+
+  for (i = 0; i < MLDSA_K; ++i)
+  __loop__(
+    assigns(MLD_IF_REDUCE_RAM(memory_slice(mat, sizeof(mld_polymat)),)
+            i, ret,
+            memory_slice(w1, sizeof(mld_poly)),
+            memory_slice(tmp, sizeof(mld_poly)),
+            memory_slice(buf, MLDSA_K * MLDSA_POLYW1_PACKEDBYTES)
+    )
+    invariant(i <= MLDSA_K)
+    decreases(MLDSA_K - i)
+  )
+  {
+    /* w1 = (A * z)_i in NTT domain */
+    mld_polyvec_matrix_pointwise_montgomery_row(w1, mat, z, i);
+
+    /* tmp = c * t1_i * 2^d in NTT domain */
+    mld_unpack_pk_t1(tmp, pk, i);
+    mld_poly_shiftl(tmp);
+    mld_poly_ntt(tmp);
+    mld_poly_pointwise_montgomery(tmp, cp);
+
+    /* w1 = invNTT(w1 - c * t1_i * 2^d) */
+    mld_poly_sub(w1, tmp);
+    mld_poly_reduce(w1);
+    mld_poly_invntt_tomont(w1);
+    mld_poly_caddq(w1);
+
+    /* tmp = h_i (decoded and validated from signature). A non-zero return
+     * means the hint encoding is malformed, i.e. the signature is invalid. */
+    if (mld_sig_unpack_hints(tmp, sig, i) != 0)
+    {
+      ret = MLD_ERR_INVALID_SIGNATURE;
+      goto cleanup;
+    }
+
+    /* w1 = use_hint(w1, tmp), then pack into buf[i] */
+    mld_poly_use_hint(w1, tmp);
+    mld_polyw1_pack(buf + i * MLDSA_POLYW1_PACKEDBYTES, w1);
+  }
+
+  /* Call random oracle and verify challenge */
+  mld_H(c2, MLDSA_CTILDEBYTES, mu, MLDSA_CRHBYTES, buf,
+        MLDSA_K * MLDSA_POLYW1_PACKEDBYTES, NULL, 0);
+
+  cmp = mld_ct_memcmp(c, c2, MLDSA_CTILDEBYTES);
+
+  /* Declassify the result of the verification. */
+  MLD_CT_TESTING_DECLASSIFY(&cmp, sizeof(cmp));
+
+  ret = cmp == 0 ? 0 : MLD_ERR_INVALID_SIGNATURE;
+
+cleanup:
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  MLD_FREE(tmp, mld_poly, 1, context);
+  MLD_FREE(w1, mld_poly, 1, context);
+  MLD_FREE(mat, mld_polymat, 1, context);
+  MLD_FREE(cp, mld_poly, 1, context);
+  MLD_FREE(z, mld_polyvecl, 1, context);
+  MLD_FREE(c2, uint8_t, MLDSA_CTILDEBYTES, context);
+  MLD_FREE(c, uint8_t, MLDSA_CTILDEBYTES, context);
+  MLD_FREE(mu, uint8_t, MLDSA_CRHBYTES, context);
+  MLD_FREE(buf, uint8_t, (MLDSA_K * MLDSA_POLYW1_PACKEDBYTES), context);
+  return ret;
+}
+
+#if !defined(MLD_CONFIG_CORE_API_ONLY)
+MLD_MUST_CHECK_RETURN_VALUE
+MLD_EXTERNAL_API
+int mld_sign_verify(const uint8_t sig[MLDSA_CRYPTO_BYTES], const uint8_t *m,
+                    size_t mlen, const uint8_t *ctx, size_t ctxlen,
+                    const uint8_t pk[MLDSA_CRYPTO_PUBLICKEYBYTES],
+                    MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  MLD_ALIGN uint8_t pre[MLD_DOMAIN_SEPARATION_MAX_BYTES];
+  size_t pre_len;
+  int ret;
+
+  pre_len = mld_prepare_domain_separation_prefix(pre, NULL, 0, ctx, ctxlen,
+                                                 MLD_PREHASH_NONE);
+  if (pre_len == 0)
+  {
+    ret = MLD_ERR_INVALID_ARG;
+    goto cleanup;
+  }
+
+  ret = mld_sign_verify_internal(sig, m, mlen, pre, pre_len, pk, 0, context);
+
+cleanup:
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  mld_zeroize(pre, sizeof(pre));
+
+  return ret;
+}
+
+MLD_MUST_CHECK_RETURN_VALUE
+MLD_EXTERNAL_API
+int mld_sign_verify_extmu(const uint8_t sig[MLDSA_CRYPTO_BYTES],
+                          const uint8_t mu[MLDSA_CRHBYTES],
+                          const uint8_t pk[MLDSA_CRYPTO_PUBLICKEYBYTES],
+                          MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  return mld_sign_verify_internal(sig, mu, MLDSA_CRHBYTES, NULL, 0, pk, 1,
+                                  context);
+}
+#endif /* !MLD_CONFIG_CORE_API_ONLY */
+#endif /* !MLD_CONFIG_NO_VERIFY_API */
+
+#if !defined(MLD_CONFIG_CORE_API_ONLY)
+#if !defined(MLD_CONFIG_NO_SIGN_API)
+MLD_MUST_CHECK_RETURN_VALUE
+MLD_EXTERNAL_API
+int mld_sign_signature_pre_hash_internal(
+    uint8_t sig[MLDSA_CRYPTO_BYTES], const uint8_t *ph, size_t phlen,
+    const uint8_t *ctx, size_t ctxlen, const uint8_t rnd[MLDSA_RNDBYTES],
+    const uint8_t sk[MLDSA_CRYPTO_SECRETKEYBYTES], int hashalg,
+    MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  MLD_ALIGN uint8_t pre[MLD_DOMAIN_SEPARATION_MAX_BYTES];
+  size_t pre_len;
+  int ret;
+
+  if (hashalg == MLD_PREHASH_NONE)
+  {
+    ret = MLD_ERR_INVALID_ARG;
+    goto cleanup;
+  }
+
+  pre_len = mld_prepare_domain_separation_prefix(pre, ph, phlen, ctx, ctxlen,
+                                                 hashalg);
+  if (pre_len == 0)
+  {
+    ret = MLD_ERR_INVALID_ARG;
+    goto cleanup;
+  }
+
+  ret = mld_sign_signature_internal(sig, pre, pre_len, NULL, 0, rnd, sk, 0,
+                                    context);
+cleanup:
+  if (ret != 0)
+  {
+    /* To be on the safe-side, make sure sig has a well-defined value, even in
+     * the case of error.
+     *
+     * If we come from mld_sign_signature_internal, this is redundant, but the
+     * error case should not be the norm, and the added cost of the zeroization
+     * insignificant. */
+    mld_zeroize(sig, MLDSA_CRYPTO_BYTES);
+  }
+
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  mld_zeroize(pre, sizeof(pre));
+  return ret;
+}
+#endif /* !MLD_CONFIG_NO_SIGN_API */
+
+#if !defined(MLD_CONFIG_NO_VERIFY_API)
+MLD_MUST_CHECK_RETURN_VALUE
+MLD_EXTERNAL_API
+int mld_sign_verify_pre_hash_internal(
+    const uint8_t sig[MLDSA_CRYPTO_BYTES], const uint8_t *ph, size_t phlen,
+    const uint8_t *ctx, size_t ctxlen,
+    const uint8_t pk[MLDSA_CRYPTO_PUBLICKEYBYTES], int hashalg,
+    MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  MLD_ALIGN uint8_t pre[MLD_DOMAIN_SEPARATION_MAX_BYTES];
+  size_t pre_len;
+  int ret;
+
+  if (hashalg == MLD_PREHASH_NONE)
+  {
+    ret = MLD_ERR_INVALID_ARG;
+    goto cleanup;
+  }
+
+  pre_len = mld_prepare_domain_separation_prefix(pre, ph, phlen, ctx, ctxlen,
+                                                 hashalg);
+  if (pre_len == 0)
+  {
+    ret = MLD_ERR_INVALID_ARG;
+    goto cleanup;
+  }
+
+  ret = mld_sign_verify_internal(sig, pre, pre_len, NULL, 0, pk, 0, context);
+
+cleanup:
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  mld_zeroize(pre, sizeof(pre));
+  return ret;
+}
+#endif /* !MLD_CONFIG_NO_VERIFY_API */
+
+#if !defined(MLD_CONFIG_NO_SIGN_API)
+MLD_MUST_CHECK_RETURN_VALUE
+MLD_EXTERNAL_API
+int mld_sign_signature_pre_hash_shake256(
+    uint8_t sig[MLDSA_CRYPTO_BYTES], const uint8_t *m, size_t mlen,
+    const uint8_t *ctx, size_t ctxlen, const uint8_t rnd[MLDSA_RNDBYTES],
+    const uint8_t sk[MLDSA_CRYPTO_SECRETKEYBYTES],
+    MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  MLD_ALIGN uint8_t ph[64];
+  int ret;
+  mld_shake256(ph, sizeof(ph), m, mlen);
+  ret = mld_sign_signature_pre_hash_internal(sig, ph, sizeof(ph), ctx, ctxlen,
+                                             rnd, sk, MLD_PREHASH_SHAKE_256,
+                                             context);
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  mld_zeroize(ph, sizeof(ph));
+  return ret;
+}
+#endif /* !MLD_CONFIG_NO_SIGN_API */
+
+#if !defined(MLD_CONFIG_NO_VERIFY_API)
+MLD_MUST_CHECK_RETURN_VALUE
+MLD_EXTERNAL_API
+int mld_sign_verify_pre_hash_shake256(
+    const uint8_t sig[MLDSA_CRYPTO_BYTES], const uint8_t *m, size_t mlen,
+    const uint8_t *ctx, size_t ctxlen,
+    const uint8_t pk[MLDSA_CRYPTO_PUBLICKEYBYTES],
+    MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  MLD_ALIGN uint8_t ph[64];
+  int ret;
+  mld_shake256(ph, sizeof(ph), m, mlen);
+  ret = mld_sign_verify_pre_hash_internal(sig, ph, sizeof(ph), ctx, ctxlen, pk,
+                                          MLD_PREHASH_SHAKE_256, context);
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  mld_zeroize(ph, sizeof(ph));
+  return ret;
+}
+#endif /* !MLD_CONFIG_NO_VERIFY_API */
+
+#if !defined(MLD_CONFIG_NO_SIGN_API) || !defined(MLD_CONFIG_NO_VERIFY_API)
+#define MLD_PRE_HASH_OID_LEN 11
+
+/**
+ * Return the OID of a given SHA-2/SHA-3 hash function.
+ *
+ * @param[out] oid     Pointer to output OID.
+ * @param      hashalg Hash algorithm constant (MLD_PREHASH_*).
+ */
+static void mld_get_hash_oid(uint8_t oid[MLD_PRE_HASH_OID_LEN], int hashalg)
+{
+  unsigned int i;
+  static const struct
+  {
+    int alg;
+    uint8_t oid[MLD_PRE_HASH_OID_LEN];
+  } oid_map[] = {
+      {MLD_PREHASH_SHA2_224,
+       {0x06, 0x09, 0x60, 0x86, 0x48, 0x01, 0x65, 0x03, 0x04, 0x02, 0x04}},
+      {MLD_PREHASH_SHA2_256,
+       {0x06, 0x09, 0x60, 0x86, 0x48, 0x01, 0x65, 0x03, 0x04, 0x02, 0x01}},
+      {MLD_PREHASH_SHA2_384,
+       {0x06, 0x09, 0x60, 0x86, 0x48, 0x01, 0x65, 0x03, 0x04, 0x02, 0x02}},
+      {MLD_PREHASH_SHA2_512,
+       {0x06, 0x09, 0x60, 0x86, 0x48, 0x01, 0x65, 0x03, 0x04, 0x02, 0x03}},
+      {MLD_PREHASH_SHA2_512_224,
+       {0x06, 0x09, 0x60, 0x86, 0x48, 0x01, 0x65, 0x03, 0x04, 0x02, 0x05}},
+      {MLD_PREHASH_SHA2_512_256,
+       {0x06, 0x09, 0x60, 0x86, 0x48, 0x01, 0x65, 0x03, 0x04, 0x02, 0x06}},
+      {MLD_PREHASH_SHA3_224,
+       {0x06, 0x09, 0x60, 0x86, 0x48, 0x01, 0x65, 0x03, 0x04, 0x02, 0x07}},
+      {MLD_PREHASH_SHA3_256,
+       {0x06, 0x09, 0x60, 0x86, 0x48, 0x01, 0x65, 0x03, 0x04, 0x02, 0x08}},
+      {MLD_PREHASH_SHA3_384,
+       {0x06, 0x09, 0x60, 0x86, 0x48, 0x01, 0x65, 0x03, 0x04, 0x02, 0x09}},
+      {MLD_PREHASH_SHA3_512,
+       {0x06, 0x09, 0x60, 0x86, 0x48, 0x01, 0x65, 0x03, 0x04, 0x02, 0x0A}},
+      {MLD_PREHASH_SHAKE_128,
+       {0x06, 0x09, 0x60, 0x86, 0x48, 0x01, 0x65, 0x03, 0x04, 0x02, 0x0B}},
+      {MLD_PREHASH_SHAKE_256,
+       {0x06, 0x09, 0x60, 0x86, 0x48, 0x01, 0x65, 0x03, 0x04, 0x02, 0x0C}}};
+
+  for (i = 0; i < sizeof(oid_map) / sizeof(oid_map[0]); i++)
+  __loop__(
+    invariant(i <= sizeof(oid_map) / sizeof(oid_map[0]))
+    decreases(sizeof(oid_map) / sizeof(oid_map[0]) - i)
+  )
+  {
+    if (oid_map[i].alg == hashalg)
+    {
+      mld_memcpy(oid, oid_map[i].oid, MLD_PRE_HASH_OID_LEN);
+      return;
+    }
+  }
+}
+
+static int mld_validate_hash_length(int hashalg, size_t len)
+{
+  switch (hashalg)
+  {
+    case MLD_PREHASH_SHA2_224:
+      return (len == 224 / 8) ? 0 : -1;
+    case MLD_PREHASH_SHA2_256:
+      return (len == 256 / 8) ? 0 : -1;
+    case MLD_PREHASH_SHA2_384:
+      return (len == 384 / 8) ? 0 : -1;
+    case MLD_PREHASH_SHA2_512:
+      return (len == 512 / 8) ? 0 : -1;
+    case MLD_PREHASH_SHA2_512_224:
+      return (len == 224 / 8) ? 0 : -1;
+    case MLD_PREHASH_SHA2_512_256:
+      return (len == 256 / 8) ? 0 : -1;
+    case MLD_PREHASH_SHA3_224:
+      return (len == 224 / 8) ? 0 : -1;
+    case MLD_PREHASH_SHA3_256:
+      return (len == 256 / 8) ? 0 : -1;
+    case MLD_PREHASH_SHA3_384:
+      return (len == 384 / 8) ? 0 : -1;
+    case MLD_PREHASH_SHA3_512:
+      return (len == 512 / 8) ? 0 : -1;
+    case MLD_PREHASH_SHAKE_128:
+      return (len == 256 / 8) ? 0 : -1;
+    case MLD_PREHASH_SHAKE_256:
+      return (len == 512 / 8) ? 0 : -1;
+    default:
+      return -1;
+  }
+}
+
+MLD_EXTERNAL_API
+size_t mld_prepare_domain_separation_prefix(
+    uint8_t prefix[MLD_DOMAIN_SEPARATION_MAX_BYTES], const uint8_t *ph,
+    size_t phlen, const uint8_t *ctx, size_t ctxlen, int hashalg)
+{
+  if (ctxlen > 255)
+  {
+    return 0;
+  }
+
+  if (hashalg != MLD_PREHASH_NONE)
+  {
+    if (ph == NULL || mld_validate_hash_length(hashalg, phlen) != 0)
+    {
+      return 0;
+    }
+  }
+
+  /* Common prefix: 0x00/0x01 || ctxlen || ctx */
+  prefix[0] = (hashalg == MLD_PREHASH_NONE) ? 0 : 1;
+  prefix[1] = (uint8_t)ctxlen;
+  if (ctxlen > 0)
+  {
+    mld_memcpy(prefix + 2, ctx, ctxlen);
+  }
+
+  if (hashalg == MLD_PREHASH_NONE)
+  {
+    return 2 + ctxlen;
+  }
+
+  /* HashML-DSA: append oid || ph */
+  mld_get_hash_oid(prefix + 2 + ctxlen, hashalg);
+  mld_memcpy(prefix + 2 + ctxlen + MLD_PRE_HASH_OID_LEN, ph, phlen);
+  return 2 + ctxlen + MLD_PRE_HASH_OID_LEN + phlen;
+}
+#endif /* !MLD_CONFIG_NO_SIGN_API || !MLD_CONFIG_NO_VERIFY_API */
+
+#if !defined(MLD_CONFIG_NO_KEYPAIR_API)
+MLD_EXTERNAL_API
+int mld_sign_pk_from_sk(uint8_t pk[MLDSA_CRYPTO_PUBLICKEYBYTES],
+                        const uint8_t sk[MLDSA_CRYPTO_SECRETKEYBYTES],
+                        MLD_CONFIG_CONTEXT_PARAMETER_TYPE context)
+{
+  uint8_t check, cmp0, cmp1, chk1, chk2;
+  int ret;
+  MLD_ALLOC(rho, uint8_t, MLDSA_SEEDBYTES, context);
+  MLD_ALLOC(tr, uint8_t, MLDSA_TRBYTES, context);
+  MLD_ALLOC(tr_computed, uint8_t, MLDSA_TRBYTES, context);
+  MLD_ALLOC(key, uint8_t, MLDSA_SEEDBYTES, context);
+  MLD_ALLOC(s1, mld_polyvecl, 1, context);
+  MLD_ALLOC(s2, mld_polyveck, 1, context);
+  MLD_ALLOC(t0_packed, uint8_t, MLDSA_K *MLDSA_POLYT0_PACKEDBYTES, context);
+
+  if (rho == NULL || tr == NULL || tr_computed == NULL || key == NULL ||
+      s1 == NULL || s2 == NULL || t0_packed == NULL)
+  {
+    ret = MLD_ERR_OUT_OF_MEMORY;
+    goto cleanup;
+  }
+
+  /* Inline unpack_sk: mld_unpack_sk uses lazy types for s1/s2/t0 which
+   * we cannot use here. t0 stays in packed form -- we compare it against
+   * the recomputed value below. */
+  mld_memcpy(rho, sk + MLDSA_SK_RHO_OFFSET, MLDSA_SEEDBYTES);
+  mld_memcpy(key, sk + MLDSA_SK_KEY_OFFSET, MLDSA_SEEDBYTES);
+  mld_memcpy(tr, sk + MLDSA_SK_TR_OFFSET, MLDSA_TRBYTES);
+  mld_polyvecl_unpack_eta(s1, sk + MLDSA_SK_S1_OFFSET);
+  mld_polyveck_unpack_eta(s2, sk + MLDSA_SK_S2_OFFSET);
+
+  /* Validate s1 and s2 coefficients are within [-MLDSA_ETA, MLDSA_ETA] */
+  chk1 = mld_polyvecl_chknorm(s1, MLDSA_ETA + 1) & 0xFF;
+  chk2 = mld_polyveck_chknorm(s2, MLDSA_ETA + 1) & 0xFF;
+
+  /* NTT s1 in place to use as s1hat */
+  mld_polyvecl_ntt(s1);
+
+  /* Pack rho into pk */
+  mld_memcpy(pk + MLDSA_PK_RHO_OFFSET, rho, MLDSA_SEEDBYTES);
+
+  /* Recompute t row by row, decompose, and pack t1 into pk and t0 into
+   * t0_packed. */
+  ret = mld_compute_pack_t0_t1(pk + MLDSA_PK_T1_OFFSET, t0_packed, s1, s2, rho,
+                               context);
+  if (ret != 0)
+  {
+    goto cleanup;
+  }
+
+  /* Compare recomputed packed t0 against the t0 region of sk. */
+  cmp0 = mld_ct_memcmp(t0_packed, sk + MLDSA_SK_T0_OFFSET,
+                       MLDSA_K * MLDSA_POLYT0_PACKEDBYTES);
+
+  /* Compute tr_computed = H(pk) and compare to the stored tr */
+  mld_shake256(tr_computed, MLDSA_TRBYTES, pk, MLDSA_CRYPTO_PUBLICKEYBYTES);
+  cmp1 = mld_ct_memcmp((const uint8_t *)tr, (const uint8_t *)tr_computed,
+                       MLDSA_TRBYTES);
+  check = mld_value_barrier_u8(cmp0 | cmp1 | chk1 | chk2);
+
+  /* Declassify the final result of the validity check. */
+  MLD_CT_TESTING_DECLASSIFY(&check, sizeof(check));
+  ret = (check != 0) ? MLD_ERR_INVALID_KEY : 0;
+
+cleanup:
+
+  if (ret != 0)
+  {
+    mld_zeroize(pk, MLDSA_CRYPTO_PUBLICKEYBYTES);
+  }
+
+  /* Constant time: pk is either the valid public key or zeroed on error */
+  MLD_CT_TESTING_DECLASSIFY(pk, MLDSA_CRYPTO_PUBLICKEYBYTES);
+
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  MLD_FREE(t0_packed, uint8_t, MLDSA_K *MLDSA_POLYT0_PACKEDBYTES, context);
+  MLD_FREE(s2, mld_polyveck, 1, context);
+  MLD_FREE(s1, mld_polyvecl, 1, context);
+  MLD_FREE(key, uint8_t, MLDSA_SEEDBYTES, context);
+  MLD_FREE(tr_computed, uint8_t, MLDSA_TRBYTES, context);
+  MLD_FREE(tr, uint8_t, MLDSA_TRBYTES, context);
+  MLD_FREE(rho, uint8_t, MLDSA_SEEDBYTES, context);
+
+  return ret;
+}
+#endif /* !MLD_CONFIG_NO_KEYPAIR_API */
+#endif /* !MLD_CONFIG_CORE_API_ONLY */
+
+/* To facilitate single-compilation-unit (SCU) builds, undefine all macros.
+ * Don't modify by hand -- this is auto-generated by scripts/autogen. */
+#undef mld_check_pct
+#undef mld_sample_s1_s2
+#undef mld_validate_hash_length
+#undef mld_get_hash_oid
+#undef mld_H
+#undef mld_compute_pack_z
+#undef mld_attempt_signature_generation
+#undef mld_compute_pack_t0_t1
+#undef mld_get_max_signing_attempts
+#undef MLD_MAX_SIGNING_ATTEMPTS
+#undef MLD_PRE_HASH_OID_LEN