s4_server/

sse.rs

//! Server-side encryption (SSE-S4) — AES-256-GCM (v0.4 #21, v0.5 #29, v0.5 #27, v0.5 #28).
//!
//! Wraps the post-compression S3 object body with authenticated
//! encryption. Compress-then-encrypt is the right order: encryption
//! produces high-entropy bytes that don't compress, so encrypting last
//! preserves the codec's ratio.
//!
//! ## Wire formats
//!
//! ### S4E1 (v0.4) — single key, no rotation
//!
//! ```text
//! [magic: "S4E1" 4B]
//! [algo:  u8]            # 1 = AES-256-GCM
//! [reserved: 3B]         # 0x00 0x00 0x00
//! [nonce: 12B]           # random per-object
//! [tag:   16B]           # AES-GCM authentication tag
//! [ciphertext: variable] # encrypted-then-authenticated body
//! ```
//!
//! Total overhead: 36 bytes per object.
//!
//! ### S4E2 (v0.5 #29) — keyring-aware, supports rotation
//!
//! ```text
//! [magic:  "S4E2" 4B]
//! [algo:   u8]            # 1 = AES-256-GCM
//! [key_id: u16 BE]        # which keyring slot encrypted this body
//! [reserved: 1B]          # 0x00
//! [nonce:  12B]           # random per-object
//! [tag:    16B]           # AES-GCM authentication tag
//! [ciphertext: variable]
//! ```
//!
//! Same 36-byte overhead — we reused the 3-byte reserved area in S4E1
//! to fit a 2-byte key-id + 1-byte reserved without bumping the header
//! size. The key-id is included in the AAD so a flipped key-id byte
//! fails the auth tag (i.e. an attacker can't trick the gateway into
//! decrypting under a different keyring slot).
//!
//! ### S4E3 (v0.5 #27) — SSE-C, customer-provided key
//!
//! ```text
//! [magic:   "S4E3" 4B]
//! [algo:    u8]            # 1 = AES-256-GCM
//! [key_md5: 16B]           # MD5 fingerprint of the customer key
//! [nonce:   12B]           # random per-object
//! [tag:     16B]           # AES-GCM authentication tag
//! [ciphertext: variable]
//! ```
//!
//! Overhead: 49 bytes (`4 + 1 + 16 + 12 + 16`). Unlike S4E1/S4E2 the
//! gateway does **not** persist the key — the client supplies it on
//! every PUT/GET via `x-amz-server-side-encryption-customer-{algorithm,
//! key,key-MD5}` headers. We store only the 16-byte MD5 in the on-disk
//! frame so a GET with the wrong key surfaces as
//! [`SseError::WrongCustomerKey`] before AES-GCM is even tried (saves a
//! useless decrypt + gives operators a distinct error from generic auth
//! failure).
//!
//! The `key_md5` is included in the AAD so flipping a single byte of
//! the stored fingerprint also breaks AES-GCM auth — i.e. an attacker
//! who tampered with the metadata can't sneak a different key past the
//! check.
//!
//! ### S4E4 (v0.5 #28) — SSE-KMS envelope, per-object DEK
//!
//! ```text
//! [magic:           "S4E4" 4B]
//! [algo:            u8]            # 1 = AES-256-GCM
//! [key_id_len:      u8]            # 1..=255, length of UTF-8 key_id
//! [key_id:          variable]      # UTF-8, AAD-authenticated
//! [wrapped_dek_len: u32 BE]        # length of the wrapped DEK blob
//! [wrapped_dek:     variable]      # opaque, AAD-authenticated
//! [nonce:           12B]           # random per-object
//! [tag:             16B]           # AES-GCM auth tag for body
//! [ciphertext:      variable]      # body encrypted under the DEK
//! ```
//!
//! Header overhead: `4 + 1 + 1 + key_id_len + 4 + wrapped_dek_len + 12
//! + 16` = 38 + key_id_len + wrapped_dek_len. For a typical
//! [`crate::kms::LocalKms`] wrap (60-byte ciphertext) and a 36-char
//! UUID-style `key_id`, that's ~134 bytes per object.
//!
//! `key_id` and `wrapped_dek` are both placed in the AAD so an
//! attacker cannot rewrite either field to point the gateway at a
//! different KEK or wrapped DEK without invalidating the body's
//! AES-GCM tag. The plaintext DEK is never persisted; only the
//! wrapped form is on disk, and the gateway holds the plaintext only
//! for the duration of one PUT or GET.
//!
//! S4E4 decrypt requires an `async` round-trip to the KMS backend
//! (to unwrap the DEK), so the synchronous [`decrypt`] function
//! refuses S4E4 with [`SseError::KmsAsyncRequired`] — callers that
//! peek `S4E4` via [`peek_magic`] must dispatch to
//! [`decrypt_with_kms`] instead.
//!
//! ## v0.5 rotation flow (SSE-S4 only)
//!
//! Operators wire one [`SseKeyring`] holding the **active** key plus
//! any number of **retired** keys. PUT always encrypts under the
//! active key (S4E2 with that key's id). GET sniffs the magic:
//!
//! - `S4E1`: legacy single-key path. The keyring's active key is tried
//!   first, then every retired key — this lets a v0.4 deployment
//!   migrate to a keyring with the original key as active and decrypt
//!   pre-rotation objects unchanged.
//! - `S4E2`: read the key_id, look it up in the keyring, decrypt with
//!   that exact key. Missing key_id surfaces as `KeyNotInKeyring`.
//! - `S4E3`: keyring is **not** consulted. Caller must supply
//!   [`SseSource::CustomerKey`] with the matching key + md5.
//!
//! ## Open follow-ups
//!
//! - **Server-managed key only** (for SSE-S4): keys come from local
//!   files via `--sse-s4-key` / `--sse-s4-key-rotated`. KMS / vault
//!   integration for the SSE-S4 keyring (i.e. wrapping the keyring's
//!   keys with KMS) is a separate issue. SSE-KMS for per-object DEKs
//!   is implemented (see [`SseSource::Kms`] + S4E4 above).

// v0.8.8: aes-gcm 0.10 + hmac 0.12 (pinned by RustCrypto) re-export the
// `Nonce::from_slice` / `Key::<Aes256Gcm>::from_slice` helpers from
// generic-array 0.14, whose helpers were deprecated in favour of the 1.x
// API. Migrating requires bumping aes-gcm to a release that pins
// generic-array 1.x (not yet stable as of this writing), so silence the
// deprecation at module scope until the upstream RustCrypto stack lands.
#![allow(deprecated)]

use std::collections::HashMap;
use std::path::Path;
use std::sync::Arc;

use aes_gcm::aead::{Aead, KeyInit, Payload};
use aes_gcm::{Aes256Gcm, Key, Nonce};
use bytes::Bytes;
use md5::{Digest as Md5Digest, Md5};
use rand::RngCore;
use thiserror::Error;

use crate::kms::{KmsBackend, KmsError, WrappedDek};

pub const SSE_MAGIC_V1: &[u8; 4] = b"S4E1";
pub const SSE_MAGIC_V2: &[u8; 4] = b"S4E2";
pub const SSE_MAGIC_V3: &[u8; 4] = b"S4E3";
pub const SSE_MAGIC_V4: &[u8; 4] = b"S4E4";
/// v0.8 #52: chunked variant of S4E2 — same SSE-S4 keyring source,
/// but the body is sliced into independently-sealed AES-GCM chunks
/// so the GET path can stream-decrypt + emit chunk-by-chunk instead
/// of buffering the entire object before tag verify. See
/// [`encrypt_v2_chunked`] / [`decrypt_chunked_stream`] for the on-
/// the-wire layout.
///
/// **Read-only as of v0.8.1 #57** — new PUTs emit [`SSE_MAGIC_V6`]
/// (S4E6). S4E5 is kept around for back-compat decrypt of objects
/// written by v0.8.0.
pub const SSE_MAGIC_V5: &[u8; 4] = b"S4E5";
/// v0.8.1 #57: identical layout to S4E5 except the per-PUT salt is
/// widened from 4 → 8 bytes so the birthday-collision threshold on
/// AES-GCM nonce reuse jumps from ~65k PUTs/key to ~4 billion. See
/// [`encrypt_v2_chunked`] (now emits S4E6) / the S4E6 wire-format
/// docs further down for the full layout.
pub const SSE_MAGIC_V6: &[u8; 4] = b"S4E6";
/// Back-compat alias — v0.4 callers that imported `SSE_MAGIC` mean S4E1.
pub const SSE_MAGIC: &[u8; 4] = SSE_MAGIC_V1;

/// Header layout matches between S4E1 and S4E2 (both 36 bytes total)
/// because S4E2 reuses the 3-byte reserved slot to fit `key_id (2B) +
/// reserved (1B)`. Keeping them the same length means the rest of the
/// pipeline (sidecar offsets, multipart math) doesn't care which
/// frame variant is in flight.
pub const SSE_HEADER_BYTES: usize = 4 + 1 + 3 + 12 + 16; // = 36
/// S4E3 (SSE-C) replaces the 3-byte reserved area with a 16-byte
/// customer-key MD5 fingerprint, so the header is 49 bytes total.
/// `magic 4 + algo 1 + key_md5 16 + nonce 12 + tag 16`.
pub const SSE_HEADER_BYTES_V3: usize = 4 + 1 + KEY_MD5_LEN + 12 + 16; // = 49
pub const ALGO_AES_256_GCM: u8 = 1;
const NONCE_LEN: usize = 12;
const TAG_LEN: usize = 16;
const KEY_LEN: usize = 32;
const KEY_MD5_LEN: usize = 16;
/// AWS S3 SSE-C only allows AES256 in the
/// `x-amz-server-side-encryption-customer-algorithm` header, so we
/// match that exact spelling for parity with real S3 clients.
pub const SSE_C_ALGORITHM: &str = "AES256";

#[derive(Debug, Error)]
pub enum SseError {
    #[error("SSE key file {path:?}: {source}")]
    KeyFileIo {
        path: std::path::PathBuf,
        source: std::io::Error,
    },
    #[error(
        "SSE key file must be exactly 32 raw bytes (or 64-char hex / 44-char base64); got {got} bytes after parse"
    )]
    BadKeyLength { got: usize },
    #[error("SSE-encrypted body too short ({got} bytes; need at least {SSE_HEADER_BYTES})")]
    TooShort { got: usize },
    #[error("SSE bad magic: expected S4E1/S4E2/S4E3/S4E4/S4E5/S4E6, got {got:?}")]
    BadMagic { got: [u8; 4] },
    #[error("SSE unsupported algo tag: {tag} (this build only knows AES-256-GCM = 1)")]
    UnsupportedAlgo { tag: u8 },
    #[error(
        "SSE key_id {id} (S4E2 frame) not present in keyring; rotation history likely incomplete"
    )]
    KeyNotInKeyring { id: u16 },
    #[error("SSE decryption / authentication failed (key mismatch or ciphertext tampered with)")]
    DecryptFailed,
    // --- v0.5 #27: SSE-C specific errors ---
    /// The MD5 fingerprint stored in the S4E3 frame doesn't match the
    /// MD5 of the customer key the client supplied. This is the
    /// "wrong customer key on GET" signal — distinct from
    /// `DecryptFailed` so service.rs can map it to AWS S3's
    /// `403 AccessDenied` (S3 returns AccessDenied when the supplied
    /// SSE-C key doesn't match the one used at PUT time).
    #[error("SSE-C key MD5 fingerprint mismatch — client supplied a different key than PUT")]
    WrongCustomerKey,
    /// `parse_customer_key_headers` saw a malformed input. `reason` is
    /// a short human string ("base64 decode of key", "key length",
    /// "md5 length", "md5 mismatch") for operator log lines — never
    /// echoed to the client (would leak crypto details).
    #[error("SSE-C customer-key headers invalid: {reason}")]
    InvalidCustomerKey { reason: &'static str },
    /// Client asked for an SSE-C algorithm the gateway doesn't speak.
    /// AWS S3 only ever defines `AES256` here; surfacing the offending
    /// string lets us 400 with a useful message.
    #[error("SSE-C algorithm {algo:?} unsupported (only {SSE_C_ALGORITHM:?} is allowed)")]
    CustomerKeyAlgorithmUnsupported { algo: String },
    /// S4E3 body lacks an SSE-C key — caller passed `SseSource::Keyring`
    /// when decrypting an SSE-C-encrypted object. service.rs should
    /// translate this into the same "missing customer key" 400 that
    /// AWS S3 returns when SSE-C headers are absent on a GET.
    #[error("S4E3 frame requires SseSource::CustomerKey; got Keyring")]
    CustomerKeyRequired,
    /// Inverse: client sent SSE-C headers on a GET for an object stored
    /// without SSE-C. The supplied key has no role in decryption, but
    /// AWS S3 actually 400s in this case ("expected an unencrypted
    /// object" / "extraneous SSE-C headers"), so we mirror that.
    #[error("S4E1/S4E2 frame stored without SSE-C; SseSource::CustomerKey is unexpected")]
    CustomerKeyUnexpected,
    // --- v0.5 #28: SSE-KMS specific errors ---
    /// `decrypt` (sync) was handed an S4E4 body. SSE-KMS unwrap is
    /// async (it round-trips to the KMS backend), so callers must
    /// peek the magic with [`peek_magic`] and dispatch S4E4 frames to
    /// [`decrypt_with_kms`] instead. service.rs's GET handler does
    /// this; tests / direct callers may hit this if they forget.
    #[error(
        "S4E4 (SSE-KMS) body requires async decrypt — call decrypt_with_kms() instead of decrypt()"
    )]
    KmsAsyncRequired,
    /// S4E4 frame is shorter than the minimum-possible header (38
    /// bytes for an empty `key_id` + empty `wrapped_dek`, which is
    /// itself impossible — we just sanity-check the floor).
    #[error("S4E4 frame too short ({got} bytes; need at least {min})")]
    KmsFrameTooShort { got: usize, min: usize },
    /// S4E4 declared a `key_id_len` or `wrapped_dek_len` that runs
    /// past the end of the body. Almost certainly truncation /
    /// corruption rather than tampering (tampering would fail the
    /// AES-GCM tag instead).
    #[error("S4E4 frame field length out of bounds: {what}")]
    KmsFrameFieldOob { what: &'static str },
    /// `key_id` field of an S4E4 frame is not valid UTF-8. We require
    /// UTF-8 because `LocalKms` uses the basename of a `.kek` file
    /// (which is OS-string-but-typically-UTF-8) and AWS KMS uses ARNs
    /// (which are ASCII).
    #[error("S4E4 key_id is not valid UTF-8")]
    KmsKeyIdNotUtf8,
    /// service.rs handed `decrypt_with_kms` a `WrappedDek` whose
    /// `key_id` doesn't match the one stored in the S4E4 frame. This
    /// is an integration bug (caller is meant to pull the wrapped
    /// DEK *from the frame*, not from somewhere else), surface as a
    /// distinct error so it shows up in tests rather than silently
    /// failing the AES-GCM tag.
    #[error(
        "S4E4 SseSource::Kms wrapped DEK key_id {supplied:?} doesn't match frame key_id {stored:?}"
    )]
    KmsWrappedDekMismatch { supplied: String, stored: String },
    /// SSE-KMS path got a non-Kms `SseSource` for an S4E4 body. The
    /// async dispatch in `decrypt_with_kms` re-derives the source
    /// internally so this can only happen if a future caller passes
    /// `SseSource::Keyring` / `CustomerKey` to a path that expected
    /// `Kms` — kept around for symmetry with the other "wrong source"
    /// errors.
    #[error("S4E4 frame requires SseSource::Kms")]
    KmsRequired,
    /// Pass-through for [`crate::kms::KmsError`] surfaced from
    /// `KmsBackend::decrypt_dek` — boxed so the variant stays small.
    #[error("KMS unwrap: {0}")]
    KmsBackend(#[from] KmsError),
    // --- v0.8 #52: S4E5 (chunked SSE-S4) specific errors ---
    /// AES-GCM auth tag verify failed on chunk `chunk_index` of an
    /// S4E5 body. Distinct from the all-or-nothing
    /// [`SseError::DecryptFailed`] because the streaming GET may
    /// have already emitted earlier chunks to the client by the
    /// time chunk N fails — operators need the chunk index in audit
    /// logs to triangulate which byte range was tampered with (or
    /// which disk sector flipped).
    #[error(
        "S4E5 chunk {chunk_index} auth tag verify failed (key mismatch or chunk tampered with)"
    )]
    ChunkAuthFailed { chunk_index: u32 },
    /// Caller asked [`encrypt_v2_chunked`] to use a chunk size of 0
    /// — nonsensical (would loop forever). Surfaced as an error
    /// rather than panicking so service.rs can map a bad
    /// `--sse-chunk-size 0` configuration to a clear startup error.
    #[error("S4E5 chunk_size must be > 0 (got 0)")]
    ChunkSizeInvalid,
    /// S4E5 frame is shorter than the fixed header or declares a
    /// (chunk_count × per-chunk-bytes) total that overruns the
    /// body. Almost certainly truncation / corruption — tampering
    /// with the per-chunk ciphertext or tag would surface as
    /// [`SseError::ChunkAuthFailed`] instead.
    #[error("S4E5 frame truncated: {what}")]
    ChunkFrameTruncated { what: &'static str },
    // --- v0.8.1 #57: S4E6 (8-byte salt, 24-bit chunk_index) ---
    /// S4E6 chunk_index is encoded as a 24-bit big-endian field in
    /// the per-chunk nonce, capping `chunk_count` at
    /// `2^24 - 1 = 16_777_215`. At the default 1 MiB chunk size that
    /// is ~16 PiB per object — well past S3's 5 GiB single-object
    /// ceiling. Surface as a distinct error so a misconfiguration
    /// (`--sse-chunk-size 1` on a multi-GiB object, say) shows up at
    /// PUT time with a clear cause rather than a panic at the u32 →
    /// u24 cast.
    #[error("S4E6 chunk_count {got} exceeds 24-bit max ({max}) — pick a larger --sse-chunk-size")]
    ChunkCountTooLarge { got: u32, max: u32 },
    // --- v0.8.2 #64: pre-allocation guard for chunked SSE frames ---
    /// `parse_chunked_header` rejected an S4E5 / S4E6 frame because
    /// the declared `chunk_size × chunk_count` (or the on-disk total
    /// after adding per-chunk tag overhead and the fixed header) is
    /// either:
    ///
    /// 1. arithmetically nonsensical (the multiplication / addition
    ///    overflows u64 on a 64-bit host), or
    /// 2. larger than the gateway's configured `max_body_bytes`
    ///    (default 5 GiB — AWS S3's single-object PUT ceiling).
    ///
    /// This is the **DoS guard** for the chunked path: without it, a
    /// 24-byte malicious header that claims `chunk_size = u32::MAX`
    /// and `chunk_count = u32::MAX` would have caused the buffered
    /// decrypt path to attempt a multi-PB `Vec::with_capacity` (or
    /// integer-overflow into a tiny alloc + later out-of-bounds
    /// panic) before any cryptographic work happened. Surface as a
    /// distinct variant — never echo the offending sizes back to the
    /// client (kept as a `&'static str` `details` field for operator
    /// audit logs only) so the response isn't a tampering oracle.
    #[error("S4E5/S4E6 chunked frame declares an over-large size: {details}")]
    ChunkFrameTooLarge { details: &'static str },
}

/// Default cap on a single chunked-SSE frame's declared plaintext
/// size, used by [`decrypt_chunked_buffered_default`] and the
/// streaming pre-validation path. Mirrors AWS S3's 5 GiB single-object
/// PUT ceiling — anything larger is unreachable on the AWS-compatible
/// API surface and is therefore safe to reject pre-alloc as a malformed
/// / malicious header (v0.8.2 #64).
pub const DEFAULT_MAX_BODY_BYTES: usize = 5 * 1024 * 1024 * 1024;

/// 32-byte symmetric key. `bytes` is `pub` so call sites can construct
/// keys directly from already-validated bytes (e.g. KMS-decrypted DEKs)
/// without going through the on-disk parser. Hold inside an `Arc` when
/// sharing across handler tasks — `SseKeyring` does this internally.
pub struct SseKey {
    pub bytes: [u8; 32],
}

impl SseKey {
    /// Load a 32-byte key from disk. Accepts three on-disk encodings:
    /// raw 32 bytes, 64-char ASCII hex, or 44-char ASCII base64 (with or
    /// without padding). Whitespace is trimmed.
    pub fn from_path(path: &Path) -> Result<Self, SseError> {
        let raw = std::fs::read(path).map_err(|source| SseError::KeyFileIo {
            path: path.to_path_buf(),
            source,
        })?;
        Self::from_bytes(&raw)
    }

    pub fn from_bytes(bytes: &[u8]) -> Result<Self, SseError> {
        // Try raw first.
        if bytes.len() == KEY_LEN {
            let mut k = [0u8; KEY_LEN];
            k.copy_from_slice(bytes);
            return Ok(Self { bytes: k });
        }
        // Trim whitespace and try hex / base64.
        let s = std::str::from_utf8(bytes).unwrap_or("").trim();
        if s.len() == KEY_LEN * 2 && s.chars().all(|c| c.is_ascii_hexdigit()) {
            let mut k = [0u8; KEY_LEN];
            for (i, k_byte) in k.iter_mut().enumerate() {
                *k_byte = u8::from_str_radix(&s[i * 2..i * 2 + 2], 16)
                    .map_err(|_| SseError::BadKeyLength { got: bytes.len() })?;
            }
            return Ok(Self { bytes: k });
        }
        if let Ok(decoded) =
            base64::Engine::decode(&base64::engine::general_purpose::STANDARD, s.as_bytes())
            && decoded.len() == KEY_LEN
        {
            let mut k = [0u8; KEY_LEN];
            k.copy_from_slice(&decoded);
            return Ok(Self { bytes: k });
        }
        Err(SseError::BadKeyLength { got: bytes.len() })
    }

    fn as_aes_key(&self) -> &Key<Aes256Gcm> {
        Key::<Aes256Gcm>::from_slice(&self.bytes)
    }
}

impl std::fmt::Debug for SseKey {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("SseKey")
            .field("len", &KEY_LEN)
            .field("key", &"<redacted>")
            .finish()
    }
}

/// v0.5 #29: a set of `SseKey`s indexed by `u16` key-id, plus a
/// designated **active** id used for new encryptions. Rotation is just
/// "add the new key, flip `active` to its id, leave the old keys for
/// decryption-only". Cheap to clone (`Arc<SseKey>` per slot).
#[derive(Clone)]
pub struct SseKeyring {
    active: u16,
    keys: HashMap<u16, Arc<SseKey>>,
}

impl SseKeyring {
    /// Create a keyring seeded with one key, immediately marked
    /// active. Add older keys later via [`SseKeyring::add`] so the
    /// gateway can still decrypt pre-rotation objects.
    pub fn new(active: u16, key: Arc<SseKey>) -> Self {
        let mut keys = HashMap::new();
        keys.insert(active, key);
        Self { active, keys }
    }

    /// Insert another key under id `id`. Does NOT change `active`. If
    /// `id == active`, the slot is overwritten (useful for tests; in
    /// production prefer minting a fresh id).
    pub fn add(&mut self, id: u16, key: Arc<SseKey>) {
        self.keys.insert(id, key);
    }

    /// Active (id, key) — used by [`encrypt_v2`] to pick the slot for
    /// new objects.
    pub fn active(&self) -> (u16, &SseKey) {
        let id = self.active;
        let key = self
            .keys
            .get(&id)
            .expect("active key id must be present in keyring (constructor invariant)");
        (id, key.as_ref())
    }

    /// Look up a key by id. Returns `None` for unknown ids — caller
    /// should surface this as [`SseError::KeyNotInKeyring`].
    pub fn get(&self, id: u16) -> Option<&SseKey> {
        self.keys.get(&id).map(Arc::as_ref)
    }
}

impl std::fmt::Debug for SseKeyring {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("SseKeyring")
            .field("active", &self.active)
            .field("key_count", &self.keys.len())
            .field("key_ids", &self.keys.keys().collect::<Vec<_>>())
            .finish()
    }
}

pub type SharedSseKeyring = Arc<SseKeyring>;

/// Encrypt `plaintext` with the given key, producing the on-the-wire
/// S4E1-framed output: `[magic 4][algo 1][reserved 3][nonce 12][tag 16][ciphertext]`.
///
/// Kept for back-compat: v0.4 callers that hand-built an `SseKey` (no
/// keyring) still get the v1 frame. New code should use
/// [`encrypt_v2`] which writes S4E2 and supports rotation on read.
pub fn encrypt(key: &SseKey, plaintext: &[u8]) -> Bytes {
    let cipher = Aes256Gcm::new(key.as_aes_key());
    let mut nonce_bytes = [0u8; NONCE_LEN];
    rand::rngs::OsRng.fill_bytes(&mut nonce_bytes);
    let nonce = Nonce::from_slice(&nonce_bytes);
    // AAD = magic + algo. Tampering with either bumps the tag check.
    let mut aad = [0u8; 8];
    aad[..4].copy_from_slice(SSE_MAGIC_V1);
    aad[4] = ALGO_AES_256_GCM;
    let ct_with_tag = cipher
        .encrypt(
            nonce,
            Payload {
                msg: plaintext,
                aad: &aad,
            },
        )
        .expect("aes-gcm encrypt cannot fail with a 32-byte key");
    debug_assert!(ct_with_tag.len() >= TAG_LEN);
    let split = ct_with_tag.len() - TAG_LEN;
    let (ct, tag) = ct_with_tag.split_at(split);

    let mut out = Vec::with_capacity(SSE_HEADER_BYTES + ct.len());
    out.extend_from_slice(SSE_MAGIC_V1);
    out.push(ALGO_AES_256_GCM);
    out.extend_from_slice(&[0u8; 3]); // reserved
    out.extend_from_slice(&nonce_bytes);
    out.extend_from_slice(tag);
    out.extend_from_slice(ct);
    Bytes::from(out)
}

/// v0.5 #29: encrypt under the keyring's currently-active key, writing
/// an S4E2-framed body (`[magic 4][algo 1][key_id 2 BE][reserved 1]
/// [nonce 12][tag 16][ciphertext]`). The key-id is included in the
/// AAD so flipping it fails the auth tag.
pub fn encrypt_v2(plaintext: &[u8], keyring: &SseKeyring) -> Bytes {
    let (key_id, key) = keyring.active();
    let cipher = Aes256Gcm::new(key.as_aes_key());
    let mut nonce_bytes = [0u8; NONCE_LEN];
    rand::rngs::OsRng.fill_bytes(&mut nonce_bytes);
    let nonce = Nonce::from_slice(&nonce_bytes);
    let aad = aad_v2(key_id);
    let ct_with_tag = cipher
        .encrypt(
            nonce,
            Payload {
                msg: plaintext,
                aad: &aad,
            },
        )
        .expect("aes-gcm encrypt cannot fail with a 32-byte key");
    debug_assert!(ct_with_tag.len() >= TAG_LEN);
    let split = ct_with_tag.len() - TAG_LEN;
    let (ct, tag) = ct_with_tag.split_at(split);

    let mut out = Vec::with_capacity(SSE_HEADER_BYTES + ct.len());
    out.extend_from_slice(SSE_MAGIC_V2);
    out.push(ALGO_AES_256_GCM);
    out.extend_from_slice(&key_id.to_be_bytes()); // 2B BE key_id
    out.push(0u8); // 1B reserved
    out.extend_from_slice(&nonce_bytes);
    out.extend_from_slice(tag);
    out.extend_from_slice(ct);
    Bytes::from(out)
}

fn aad_v1() -> [u8; 8] {
    let mut aad = [0u8; 8];
    aad[..4].copy_from_slice(SSE_MAGIC_V1);
    aad[4] = ALGO_AES_256_GCM;
    aad
}

fn aad_v2(key_id: u16) -> [u8; 8] {
    let mut aad = [0u8; 8];
    aad[..4].copy_from_slice(SSE_MAGIC_V2);
    aad[4] = ALGO_AES_256_GCM;
    aad[5..7].copy_from_slice(&key_id.to_be_bytes());
    aad[7] = 0u8;
    aad
}

/// AAD for S4E3 = magic (4) + algo (1) + key_md5 (16). Putting the
/// fingerprint in the AAD means tampering with the stored MD5 (e.g. an
/// attacker rewriting the header to match a *different* key they
/// happen to know) breaks the AES-GCM tag — the wrong-key check isn't
/// just a plain `==` we could be tricked past.
fn aad_v3(key_md5: &[u8; KEY_MD5_LEN]) -> [u8; 4 + 1 + KEY_MD5_LEN] {
    let mut aad = [0u8; 4 + 1 + KEY_MD5_LEN];
    aad[..4].copy_from_slice(SSE_MAGIC_V3);
    aad[4] = ALGO_AES_256_GCM;
    aad[5..5 + KEY_MD5_LEN].copy_from_slice(key_md5);
    aad
}

/// Parsed + verified SSE-C key material from the three customer
/// headers. `key_md5` is the MD5 of `key` (we recompute and compare in
/// [`parse_customer_key_headers`] — clients send their own to catch
/// transport corruption, but we *trust* our own computation as the
/// canonical fingerprint in the S4E3 frame).
#[derive(Clone)]
pub struct CustomerKeyMaterial {
    pub key: [u8; KEY_LEN],
    pub key_md5: [u8; KEY_MD5_LEN],
}

impl std::fmt::Debug for CustomerKeyMaterial {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        // Don't leak the key into logs. The MD5 is a public fingerprint
        // (S3 puts it on the wire), so that's safe to show.
        f.debug_struct("CustomerKeyMaterial")
            .field("key", &"<redacted>")
            .field("key_md5_hex", &hex_lower(&self.key_md5))
            .finish()
    }
}

fn hex_lower(bytes: &[u8]) -> String {
    let mut s = String::with_capacity(bytes.len() * 2);
    for b in bytes {
        s.push_str(&format!("{b:02x}"));
    }
    s
}

/// Source of the encryption key for [`encrypt_with_source`] /
/// [`decrypt`]. SSE-S4 (server-managed, rotation-aware) goes through
/// `Keyring`; SSE-C (customer-supplied) goes through `CustomerKey`.
///
/// Borrowed (not owned) so the caller can hold a long-lived
/// `CustomerKeyMaterial` next to the request and just lend it for the
/// duration of one PUT/GET.
#[derive(Debug, Clone, Copy)]
pub enum SseSource<'a> {
    /// Server-managed keyring path → produces / consumes S4E1 (legacy)
    /// or S4E2 (rotation-aware) frames.
    Keyring(&'a SseKeyring),
    /// Client-supplied AES-256 key + its MD5 fingerprint → produces /
    /// consumes S4E3 frames. The server never persists the key; it
    /// stores `key_md5` only.
    CustomerKey {
        key: &'a [u8; KEY_LEN],
        key_md5: &'a [u8; KEY_MD5_LEN],
    },
    /// SSE-KMS envelope → produces / consumes S4E4 frames. The server
    /// holds a per-object plaintext DEK (from a fresh
    /// [`KmsBackend::generate_dek`] call) and the wrapped form to
    /// persist alongside the body. The DEK is dropped after one
    /// PUT/GET; only the wrapped form survives at rest.
    Kms {
        /// 32-byte plaintext DEK, used as the AES-GCM key.
        dek: &'a [u8; KEY_LEN],
        /// Wrapped form to persist in the S4E4 frame (PUT) or the one
        /// read out of the frame (GET, after a successful unwrap).
        wrapped: &'a WrappedDek,
    },
}

/// Back-compat coercion: existing call sites pass `&SseKeyring`
/// directly to [`decrypt`]. With this `From` impl the generic bound
/// `Into<SseSource>` accepts `&SseKeyring` without the caller writing
/// `.into()`, keeping v0.4 / v0.5 #29 service.rs callers compiling
/// untouched while v0.5 #27 SSE-C callers pass `SseSource::CustomerKey`
/// explicitly.
impl<'a> From<&'a SseKeyring> for SseSource<'a> {
    fn from(kr: &'a SseKeyring) -> Self {
        SseSource::Keyring(kr)
    }
}

/// service.rs holds keyring as `Option<Arc<SseKeyring>>` and unwraps to
/// `&Arc<SseKeyring>` — let that coerce too, otherwise every existing
/// call site needs `.as_ref()` boilerplate.
impl<'a> From<&'a Arc<SseKeyring>> for SseSource<'a> {
    fn from(kr: &'a Arc<SseKeyring>) -> Self {
        SseSource::Keyring(kr.as_ref())
    }
}

impl<'a> From<&'a CustomerKeyMaterial> for SseSource<'a> {
    fn from(m: &'a CustomerKeyMaterial) -> Self {
        SseSource::CustomerKey {
            key: &m.key,
            key_md5: &m.key_md5,
        }
    }
}

/// Parse the three AWS SSE-C headers and return verified key material.
///
/// Validates, in order:
/// 1. `algorithm == "AES256"` (the only value AWS S3 defines).
/// 2. `key_base64` decodes to exactly 32 bytes (AES-256 key length).
/// 3. `key_md5_base64` decodes to exactly 16 bytes (MD5 digest length).
/// 4. The actual MD5 of the decoded key matches the supplied MD5.
///
/// Step 4 catches transport corruption *and* a class of programming
/// bugs where the client signs with one key but uploads another. AWS
/// S3 also performs this check.
pub fn parse_customer_key_headers(
    algorithm: &str,
    key_base64: &str,
    key_md5_base64: &str,
) -> Result<CustomerKeyMaterial, SseError> {
    use base64::Engine as _;
    if algorithm != SSE_C_ALGORITHM {
        return Err(SseError::CustomerKeyAlgorithmUnsupported {
            algo: algorithm.to_string(),
        });
    }
    let key_bytes = base64::engine::general_purpose::STANDARD
        .decode(key_base64.trim().as_bytes())
        .map_err(|_| SseError::InvalidCustomerKey {
            reason: "base64 decode of key",
        })?;
    if key_bytes.len() != KEY_LEN {
        return Err(SseError::InvalidCustomerKey {
            reason: "key length (must be 32 bytes after base64 decode)",
        });
    }
    let supplied_md5 = base64::engine::general_purpose::STANDARD
        .decode(key_md5_base64.trim().as_bytes())
        .map_err(|_| SseError::InvalidCustomerKey {
            reason: "base64 decode of key MD5",
        })?;
    if supplied_md5.len() != KEY_MD5_LEN {
        return Err(SseError::InvalidCustomerKey {
            reason: "key MD5 length (must be 16 bytes after base64 decode)",
        });
    }
    let actual_md5 = compute_key_md5(&key_bytes);
    // Constant-time compare — paranoia, MD5 is non-secret but the key
    // it identifies is, so we don't want a timing oracle.
    if !constant_time_eq(&actual_md5, &supplied_md5) {
        return Err(SseError::InvalidCustomerKey {
            reason: "supplied MD5 does not match MD5 of supplied key",
        });
    }
    let mut key = [0u8; KEY_LEN];
    key.copy_from_slice(&key_bytes);
    let mut key_md5 = [0u8; KEY_MD5_LEN];
    key_md5.copy_from_slice(&actual_md5);
    Ok(CustomerKeyMaterial { key, key_md5 })
}

/// Convenience wrapper — compute the MD5 fingerprint of a 32-byte
/// customer key. Callers that already have the bytes (e.g. derived
/// from a KMS unwrap) can use this to construct a
/// [`CustomerKeyMaterial`] directly.
pub fn compute_key_md5(key: &[u8]) -> [u8; KEY_MD5_LEN] {
    let mut h = Md5::new();
    h.update(key);
    let out = h.finalize();
    let mut md5 = [0u8; KEY_MD5_LEN];
    md5.copy_from_slice(&out);
    md5
}

/// `subtle`-free constant-time byte slice equality. We only need this
/// at one site (MD5 verification) so pulling `subtle` in feels excessive.
fn constant_time_eq(a: &[u8], b: &[u8]) -> bool {
    if a.len() != b.len() {
        return false;
    }
    let mut acc: u8 = 0;
    for (x, y) in a.iter().zip(b.iter()) {
        acc |= x ^ y;
    }
    acc == 0
}

/// v0.5 #27: encrypt under whichever source the caller picked.
///
/// - `SseSource::Keyring` → delegates to [`encrypt_v2`] (S4E2 frame).
/// - `SseSource::CustomerKey` → writes an S4E3 frame (no key persisted,
///   just the MD5 fingerprint for GET-side verification).
///
/// service.rs picks the source per-request: SSE-C headers present →
/// `CustomerKey`, otherwise (and only when `--sse-s4-key` is wired) →
/// `Keyring`. Plaintext objects skip this function entirely.
pub fn encrypt_with_source(plaintext: &[u8], source: SseSource<'_>) -> Bytes {
    match source {
        SseSource::Keyring(kr) => encrypt_v2(plaintext, kr),
        SseSource::CustomerKey { key, key_md5 } => encrypt_v3(plaintext, key, key_md5),
        SseSource::Kms { dek, wrapped } => encrypt_v4(plaintext, dek, wrapped),
    }
}

fn encrypt_v3(plaintext: &[u8], key: &[u8; KEY_LEN], key_md5: &[u8; KEY_MD5_LEN]) -> Bytes {
    let aes_key = Key::<Aes256Gcm>::from_slice(key);
    let cipher = Aes256Gcm::new(aes_key);
    let mut nonce_bytes = [0u8; NONCE_LEN];
    rand::rngs::OsRng.fill_bytes(&mut nonce_bytes);
    let nonce = Nonce::from_slice(&nonce_bytes);
    let aad = aad_v3(key_md5);
    let ct_with_tag = cipher
        .encrypt(
            nonce,
            Payload {
                msg: plaintext,
                aad: &aad,
            },
        )
        .expect("aes-gcm encrypt cannot fail with a 32-byte key");
    debug_assert!(ct_with_tag.len() >= TAG_LEN);
    let split = ct_with_tag.len() - TAG_LEN;
    let (ct, tag) = ct_with_tag.split_at(split);

    let mut out = Vec::with_capacity(SSE_HEADER_BYTES_V3 + ct.len());
    out.extend_from_slice(SSE_MAGIC_V3);
    out.push(ALGO_AES_256_GCM);
    out.extend_from_slice(key_md5);
    out.extend_from_slice(&nonce_bytes);
    out.extend_from_slice(tag);
    out.extend_from_slice(ct);
    Bytes::from(out)
}

/// v0.5 #29 + v0.5 #27: dispatch on the body's magic and decrypt under
/// whichever source the caller supplied.
///
/// - `S4E1` / `S4E2` require `SseSource::Keyring` (return
///   [`SseError::CustomerKeyRequired`] for `CustomerKey` — service.rs
///   should map this to "extraneous SSE-C headers" 400).
/// - `S4E3` requires `SseSource::CustomerKey` (return
///   [`SseError::CustomerKeyUnexpected`] for `Keyring` — service.rs
///   should map this to "missing SSE-C headers" 400).
///
/// Generic over `Into<SseSource>` so existing `decrypt(body, &keyring)`
/// call sites compile unchanged via the `From<&SseKeyring>` impl above
/// — only the new SSE-C path needs to type out
/// `SseSource::CustomerKey { .. }`.
///
/// Distinct errors (`KeyNotInKeyring`, `DecryptFailed`,
/// `WrongCustomerKey`) let operators tell rotation gaps, ciphertext
/// tampering, and SSE-C key mismatch apart in audit logs.
pub fn decrypt<'a, S: Into<SseSource<'a>>>(body: &[u8], source: S) -> Result<Bytes, SseError> {
    let source = source.into();
    // Outer short-check uses the smaller of the two header sizes
    // (S4E1/S4E2 = 36 bytes). Anything below this can't be any valid
    // SSE frame regardless of magic — keeps back-compat with v0.4 /
    // v0.5 #29 callers that expected `TooShort` for absurdly short
    // inputs even when the magic is garbage.
    if body.len() < SSE_HEADER_BYTES {
        return Err(SseError::TooShort { got: body.len() });
    }
    let mut magic = [0u8; 4];
    magic.copy_from_slice(&body[..4]);
    match &magic {
        m if m == SSE_MAGIC_V1 || m == SSE_MAGIC_V2 => {
            let keyring = match source {
                SseSource::Keyring(kr) => kr,
                SseSource::CustomerKey { .. } => return Err(SseError::CustomerKeyUnexpected),
                // S4E1/E2 stored under the keyring → SseSource::Kms
                // is just as nonsensical as CustomerKey here. Re-use
                // the same "wrong source" error so service.rs can
                // map both to AWS S3's "extraneous SSE-* headers"
                // 400.
                SseSource::Kms { .. } => return Err(SseError::CustomerKeyUnexpected),
            };
            if m == SSE_MAGIC_V1 {
                decrypt_v1_with_keyring(body, keyring)
            } else {
                decrypt_v2_with_keyring(body, keyring)
            }
        }
        m if m == SSE_MAGIC_V3 => {
            // S4E3 has a larger 49-byte header, so re-check.
            if body.len() < SSE_HEADER_BYTES_V3 {
                return Err(SseError::TooShort { got: body.len() });
            }
            let (key, key_md5) = match source {
                SseSource::CustomerKey { key, key_md5 } => (key, key_md5),
                SseSource::Keyring(_) => return Err(SseError::CustomerKeyRequired),
                SseSource::Kms { .. } => return Err(SseError::CustomerKeyRequired),
            };
            decrypt_v3(body, key, key_md5)
        }
        m if m == SSE_MAGIC_V4 => {
            // SSE-KMS unwrap is async (KMS round-trip required).
            // Caller must dispatch to `decrypt_with_kms` after
            // peeking the magic — surface this as a distinct error
            // rather than silently failing.
            Err(SseError::KmsAsyncRequired)
        }
        m if m == SSE_MAGIC_V5 || m == SSE_MAGIC_V6 => {
            // v0.8 #52 (S4E5) / v0.8.1 #57 (S4E6): chunked SSE-S4.
            // Sync back-compat path — verifies + decrypts every
            // chunk into a single Bytes. Callers that want true
            // streaming (per-chunk emit) should use
            // `decrypt_chunked_stream` instead. SSE-C and SSE-KMS
            // sources are nonsensical here for the same reason as
            // S4E2 (server-managed keyring only).
            let keyring = match source {
                SseSource::Keyring(kr) => kr,
                SseSource::CustomerKey { .. } => {
                    return Err(SseError::CustomerKeyUnexpected);
                }
                SseSource::Kms { .. } => return Err(SseError::CustomerKeyUnexpected),
            };
            // v0.8.2 #64: route through the back-compat wrapper so
            // the public `decrypt` signature stays stable while the
            // pre-allocation guard runs against the gateway's
            // 5 GiB single-object cap. Bespoke callers (e.g. tests
            // or future API surfaces that want a tighter limit) can
            // call `decrypt_chunked_buffered` directly.
            decrypt_chunked_buffered_default(body, keyring)
        }
        _ => Err(SseError::BadMagic { got: magic }),
    }
}

fn decrypt_v3(
    body: &[u8],
    key: &[u8; KEY_LEN],
    supplied_md5: &[u8; KEY_MD5_LEN],
) -> Result<Bytes, SseError> {
    let algo = body[4];
    if algo != ALGO_AES_256_GCM {
        return Err(SseError::UnsupportedAlgo { tag: algo });
    }
    let mut stored_md5 = [0u8; KEY_MD5_LEN];
    stored_md5.copy_from_slice(&body[5..5 + KEY_MD5_LEN]);
    // Cheap fingerprint check first — if the supplied key has a
    // different MD5 than what was used at PUT, fail fast with a
    // dedicated error. AES-GCM auth would also catch this (different
    // key → bad tag) but the bespoke error gives operators an audit
    // signal distinct from "ciphertext was tampered with".
    if !constant_time_eq(supplied_md5, &stored_md5) {
        return Err(SseError::WrongCustomerKey);
    }
    let nonce_off = 5 + KEY_MD5_LEN;
    let tag_off = nonce_off + NONCE_LEN;
    let mut nonce_bytes = [0u8; NONCE_LEN];
    nonce_bytes.copy_from_slice(&body[nonce_off..nonce_off + NONCE_LEN]);
    let mut tag_bytes = [0u8; TAG_LEN];
    tag_bytes.copy_from_slice(&body[tag_off..tag_off + TAG_LEN]);
    let ct = &body[SSE_HEADER_BYTES_V3..];

    let aad = aad_v3(&stored_md5);
    let nonce = Nonce::from_slice(&nonce_bytes);
    let mut ct_with_tag = Vec::with_capacity(ct.len() + TAG_LEN);
    ct_with_tag.extend_from_slice(ct);
    ct_with_tag.extend_from_slice(&tag_bytes);

    let aes_key = Key::<Aes256Gcm>::from_slice(key);
    let cipher = Aes256Gcm::new(aes_key);
    let plain = cipher
        .decrypt(
            nonce,
            Payload {
                msg: &ct_with_tag,
                aad: &aad,
            },
        )
        .map_err(|_| SseError::DecryptFailed)?;
    Ok(Bytes::from(plain))
}

/// AAD for S4E4 = magic (4) + algo (1) + key_id_len (1) + key_id +
/// wrapped_dek_len (4 BE) + wrapped_dek. Putting the variable-length
/// key_id and wrapped_dek into the AAD means an attacker cannot
/// rewrite either field to redirect the gateway to a different KEK
/// or wrapped DEK without invalidating the body's AES-GCM tag.
///
/// Length-prefixing key_id and wrapped_dek inside the AAD prevents a
/// canonicalisation ambiguity: without the length prefix, an
/// attacker could shift bytes between the two fields and produce the
/// same AAD bytestream, defeating the per-field tampering check.
fn aad_v4(key_id: &[u8], wrapped_dek: &[u8]) -> Vec<u8> {
    let mut aad = Vec::with_capacity(4 + 1 + 1 + key_id.len() + 4 + wrapped_dek.len());
    aad.extend_from_slice(SSE_MAGIC_V4);
    aad.push(ALGO_AES_256_GCM);
    aad.push(key_id.len() as u8);
    aad.extend_from_slice(key_id);
    aad.extend_from_slice(&(wrapped_dek.len() as u32).to_be_bytes());
    aad.extend_from_slice(wrapped_dek);
    aad
}

fn encrypt_v4(plaintext: &[u8], dek: &[u8; KEY_LEN], wrapped: &WrappedDek) -> Bytes {
    // Pre-conditions: key_id must fit in a u8 length prefix and be
    // non-empty (an empty id means we wouldn't be able to re-fetch
    // the KEK on GET). wrapped_dek length fits in u32 by the same
    // logic — at u32::MAX bytes you have bigger problems. We assert
    // these in debug and silently truncate-or-panic in release; in
    // practice key_id is a UUID or ARN (<256 chars) and wrapped_dek
    // is 60 bytes (LocalKms) or ~200 bytes (AWS KMS).
    assert!(
        !wrapped.key_id.is_empty() && wrapped.key_id.len() <= u8::MAX as usize,
        "S4E4 key_id must be 1..=255 bytes (got {})",
        wrapped.key_id.len()
    );
    assert!(
        wrapped.ciphertext.len() <= u32::MAX as usize,
        "S4E4 wrapped_dek longer than u32::MAX",
    );

    let aes_key = Key::<Aes256Gcm>::from_slice(dek);
    let cipher = Aes256Gcm::new(aes_key);
    let mut nonce_bytes = [0u8; NONCE_LEN];
    rand::rngs::OsRng.fill_bytes(&mut nonce_bytes);
    let nonce = Nonce::from_slice(&nonce_bytes);
    let aad = aad_v4(wrapped.key_id.as_bytes(), &wrapped.ciphertext);
    let ct_with_tag = cipher
        .encrypt(
            nonce,
            Payload {
                msg: plaintext,
                aad: &aad,
            },
        )
        .expect("aes-gcm encrypt cannot fail with a 32-byte key");
    debug_assert!(ct_with_tag.len() >= TAG_LEN);
    let split = ct_with_tag.len() - TAG_LEN;
    let (ct, tag) = ct_with_tag.split_at(split);

    let key_id_bytes = wrapped.key_id.as_bytes();
    let mut out = Vec::with_capacity(
        4 + 1
            + 1
            + key_id_bytes.len()
            + 4
            + wrapped.ciphertext.len()
            + NONCE_LEN
            + TAG_LEN
            + ct.len(),
    );
    out.extend_from_slice(SSE_MAGIC_V4);
    out.push(ALGO_AES_256_GCM);
    out.push(key_id_bytes.len() as u8);
    out.extend_from_slice(key_id_bytes);
    out.extend_from_slice(&(wrapped.ciphertext.len() as u32).to_be_bytes());
    out.extend_from_slice(&wrapped.ciphertext);
    out.extend_from_slice(&nonce_bytes);
    out.extend_from_slice(tag);
    out.extend_from_slice(ct);
    Bytes::from(out)
}

/// Parsed view of an S4E4 frame's variable-length header. Returned
/// by [`parse_s4e4_header`] so both the async [`decrypt_with_kms`]
/// path and any future inspection code (e.g. an admin tool that
/// needs to enumerate object → KMS-key bindings) can reuse the same
/// parser without re-implementing offset math.
#[derive(Debug)]
pub struct S4E4Header<'a> {
    pub key_id: &'a str,
    pub wrapped_dek: &'a [u8],
    pub nonce: &'a [u8],
    pub tag: &'a [u8],
    pub ciphertext: &'a [u8],
}

/// Parse the (variable-length) S4E4 header. Pure byte-shuffling — no
/// crypto, no KMS round-trip. Returns errors on truncation /
/// out-of-bounds field lengths / non-UTF-8 key_id.
pub fn parse_s4e4_header(body: &[u8]) -> Result<S4E4Header<'_>, SseError> {
    // Minimum: magic(4) + algo(1) + key_id_len(1) + key_id(>=1) +
    // wrapped_dek_len(4) + wrapped_dek(>=1) + nonce(12) + tag(16)
    // = 40 bytes. We use a slightly looser floor here (bytes for
    // empty fields = 38) and let the per-field bounds checks below
    // catch the actual short reads.
    const S4E4_MIN: usize = 4 + 1 + 1 + 4 + NONCE_LEN + TAG_LEN; // 38
    if body.len() < S4E4_MIN {
        return Err(SseError::KmsFrameTooShort {
            got: body.len(),
            min: S4E4_MIN,
        });
    }
    let magic = &body[..4];
    if magic != SSE_MAGIC_V4 {
        let mut got = [0u8; 4];
        got.copy_from_slice(magic);
        return Err(SseError::BadMagic { got });
    }
    let algo = body[4];
    if algo != ALGO_AES_256_GCM {
        return Err(SseError::UnsupportedAlgo { tag: algo });
    }
    let key_id_len = body[5] as usize;
    let key_id_off: usize = 6;
    let key_id_end = key_id_off
        .checked_add(key_id_len)
        .ok_or(SseError::KmsFrameFieldOob { what: "key_id_len" })?;
    if key_id_end + 4 > body.len() {
        return Err(SseError::KmsFrameFieldOob { what: "key_id" });
    }
    let key_id = std::str::from_utf8(&body[key_id_off..key_id_end])
        .map_err(|_| SseError::KmsKeyIdNotUtf8)?;
    let wrapped_len_off = key_id_end;
    let wrapped_dek_len = u32::from_be_bytes([
        body[wrapped_len_off],
        body[wrapped_len_off + 1],
        body[wrapped_len_off + 2],
        body[wrapped_len_off + 3],
    ]) as usize;
    let wrapped_off = wrapped_len_off + 4;
    let wrapped_end =
        wrapped_off
            .checked_add(wrapped_dek_len)
            .ok_or(SseError::KmsFrameFieldOob {
                what: "wrapped_dek_len",
            })?;
    if wrapped_end + NONCE_LEN + TAG_LEN > body.len() {
        return Err(SseError::KmsFrameFieldOob {
            what: "wrapped_dek",
        });
    }
    let wrapped_dek = &body[wrapped_off..wrapped_end];
    let nonce_off = wrapped_end;
    let tag_off = nonce_off + NONCE_LEN;
    let ct_off = tag_off + TAG_LEN;
    let nonce = &body[nonce_off..nonce_off + NONCE_LEN];
    let tag = &body[tag_off..tag_off + TAG_LEN];
    let ciphertext = &body[ct_off..];
    Ok(S4E4Header {
        key_id,
        wrapped_dek,
        nonce,
        tag,
        ciphertext,
    })
}

/// Async decrypt for S4E4 (SSE-KMS) bodies. Caller supplies the KMS
/// backend; this function parses the frame, calls
/// `kms.decrypt_dek(...)` to unwrap the DEK, then runs AES-256-GCM
/// to recover the plaintext.
///
/// service.rs's GET handler should peek the magic with [`peek_magic`]
/// and dispatch:
///
/// - `Some("S4E4")` → `decrypt_with_kms(blob, &*kms).await`
/// - everything else → existing sync `decrypt(blob, source)`
///
/// Note: we don't go through `SseSource::Kms` here because the
/// wrapped DEK + key_id come from the frame itself, not from the
/// request — the `SseSource` is built for sync paths where the
/// caller already knows the key.
pub async fn decrypt_with_kms(body: &[u8], kms: &dyn KmsBackend) -> Result<Bytes, SseError> {
    let hdr = parse_s4e4_header(body)?;
    let wrapped = WrappedDek {
        key_id: hdr.key_id.to_string(),
        ciphertext: hdr.wrapped_dek.to_vec(),
    };
    let dek_vec = kms.decrypt_dek(&wrapped).await?;
    if dek_vec.len() != KEY_LEN {
        // KMS returned a non-32-byte plaintext. AES-256 needs exactly
        // 32 bytes. This shouldn't happen with `KeySpec=AES_256` but
        // surface as a backend error so it's auditable rather than
        // panicking.
        return Err(SseError::KmsBackend(KmsError::BackendUnavailable {
            message: format!(
                "KMS returned {} byte DEK; expected {KEY_LEN}",
                dek_vec.len()
            ),
        }));
    }
    let mut dek = [0u8; KEY_LEN];
    dek.copy_from_slice(&dek_vec);

    let aad = aad_v4(hdr.key_id.as_bytes(), hdr.wrapped_dek);
    let aes_key = Key::<Aes256Gcm>::from_slice(&dek);
    let cipher = Aes256Gcm::new(aes_key);
    let nonce = Nonce::from_slice(hdr.nonce);
    let mut ct_with_tag = Vec::with_capacity(hdr.ciphertext.len() + TAG_LEN);
    ct_with_tag.extend_from_slice(hdr.ciphertext);
    ct_with_tag.extend_from_slice(hdr.tag);
    let plain = cipher
        .decrypt(
            nonce,
            Payload {
                msg: &ct_with_tag,
                aad: &aad,
            },
        )
        .map_err(|_| SseError::DecryptFailed)?;
    Ok(Bytes::from(plain))
}

fn decrypt_v1_with_keyring(body: &[u8], keyring: &SseKeyring) -> Result<Bytes, SseError> {
    let algo = body[4];
    if algo != ALGO_AES_256_GCM {
        return Err(SseError::UnsupportedAlgo { tag: algo });
    }
    // body[5..8] reserved (must be ignored — v0.4 wrote zeros, but we
    // didn't auth them so we can't insist on it).
    let mut nonce_bytes = [0u8; NONCE_LEN];
    nonce_bytes.copy_from_slice(&body[8..8 + NONCE_LEN]);
    let mut tag_bytes = [0u8; TAG_LEN];
    tag_bytes.copy_from_slice(&body[8 + NONCE_LEN..SSE_HEADER_BYTES]);
    let ct = &body[SSE_HEADER_BYTES..];

    let aad = aad_v1();
    let nonce = Nonce::from_slice(&nonce_bytes);
    let mut ct_with_tag = Vec::with_capacity(ct.len() + TAG_LEN);
    ct_with_tag.extend_from_slice(ct);
    ct_with_tag.extend_from_slice(&tag_bytes);

    // Active key first, then any others. v0.4 deployments that flip to
    // v0.5 with their original key as active hit this path on the
    // first try for every legacy object.
    let (active_id, _active_key) = keyring.active();
    let mut ids: Vec<u16> = keyring.keys.keys().copied().collect();
    ids.sort_by_key(|id| if *id == active_id { 0 } else { 1 });
    for id in ids {
        let key = keyring.get(id).expect("id came from keyring iteration");
        let cipher = Aes256Gcm::new(key.as_aes_key());
        if let Ok(plain) = cipher.decrypt(
            nonce,
            Payload {
                msg: &ct_with_tag,
                aad: &aad,
            },
        ) {
            return Ok(Bytes::from(plain));
        }
    }
    Err(SseError::DecryptFailed)
}

fn decrypt_v2_with_keyring(body: &[u8], keyring: &SseKeyring) -> Result<Bytes, SseError> {
    let algo = body[4];
    if algo != ALGO_AES_256_GCM {
        return Err(SseError::UnsupportedAlgo { tag: algo });
    }
    let key_id = u16::from_be_bytes([body[5], body[6]]);
    // body[7] reserved (1B), authenticated as 0 via AAD.
    let key = keyring
        .get(key_id)
        .ok_or(SseError::KeyNotInKeyring { id: key_id })?;
    let mut nonce_bytes = [0u8; NONCE_LEN];
    nonce_bytes.copy_from_slice(&body[8..8 + NONCE_LEN]);
    let mut tag_bytes = [0u8; TAG_LEN];
    tag_bytes.copy_from_slice(&body[8 + NONCE_LEN..SSE_HEADER_BYTES]);
    let ct = &body[SSE_HEADER_BYTES..];

    let aad = aad_v2(key_id);
    let nonce = Nonce::from_slice(&nonce_bytes);
    let mut ct_with_tag = Vec::with_capacity(ct.len() + TAG_LEN);
    ct_with_tag.extend_from_slice(ct);
    ct_with_tag.extend_from_slice(&tag_bytes);
    let cipher = Aes256Gcm::new(key.as_aes_key());
    let plain = cipher
        .decrypt(
            nonce,
            Payload {
                msg: &ct_with_tag,
                aad: &aad,
            },
        )
        .map_err(|_| SseError::DecryptFailed)?;
    Ok(Bytes::from(plain))
}

/// Detect whether `body` is SSE-S4 encrypted (S4E1, S4E2, S4E3, or
/// S4E4) by sniffing the first 4 magic bytes. Used by the GET path
/// to decide whether to run decryption before frame parsing.
///
/// We require a length check that's safe for *any* of the four
/// frames — `SSE_HEADER_BYTES` (36) is the smallest valid header
/// (S4E1 / S4E2). S4E3 is 49 bytes; S4E4 is variable but always >=
/// 38 bytes. The per-frame decrypt path re-checks the appropriate
/// minimum, so this 36-byte gate is just a fast rejection of
/// obviously-too-short bodies.
pub fn looks_encrypted(body: &[u8]) -> bool {
    if body.len() < SSE_HEADER_BYTES {
        return false;
    }
    let m = &body[..4];
    m == SSE_MAGIC_V1
        || m == SSE_MAGIC_V2
        || m == SSE_MAGIC_V3
        || m == SSE_MAGIC_V4
        || m == SSE_MAGIC_V5
        || m == SSE_MAGIC_V6
}

/// Peek the SSE-S4 magic at the front of `body`, returning a
/// stringified frame variant identifier or `None` if `body` is not
/// recognized as SSE-S4. Used by the GET path to dispatch between
/// the sync [`decrypt`] (S4E1/E2/E3) and the async
/// [`decrypt_with_kms`] (S4E4).
///
/// Returns the same length-gated result as [`looks_encrypted`]: any
/// body shorter than `SSE_HEADER_BYTES` (36 bytes) returns `None`,
/// so the caller can use this as both the "is encrypted" signal and
/// the "which frame" signal in one cheap byte-comparison.
pub fn peek_magic(body: &[u8]) -> Option<&'static str> {
    if body.len() < SSE_HEADER_BYTES {
        return None;
    }
    match &body[..4] {
        m if m == SSE_MAGIC_V1 => Some("S4E1"),
        m if m == SSE_MAGIC_V2 => Some("S4E2"),
        m if m == SSE_MAGIC_V3 => Some("S4E3"),
        m if m == SSE_MAGIC_V4 => Some("S4E4"),
        // v0.8 #52: chunked SSE-S4. service.rs's GET handler
        // dispatches "S4E5" / "S4E6" → `decrypt_chunked_stream`
        // for true streaming GET; the sync `decrypt(...)` also
        // accepts both (back-compat — buffered concat).
        m if m == SSE_MAGIC_V5 => Some("S4E5"),
        // v0.8.1 #57: same dispatch as S4E5 — wider salt only.
        m if m == SSE_MAGIC_V6 => Some("S4E6"),
        _ => None,
    }
}

pub type SharedSseKey = Arc<SseKey>;

// ===========================================================================
// v0.8 #52 (S4E5, read-only) + v0.8.1 #57 (S4E6, current emit) —
// chunked variant of S4E2 for streaming GET
// ===========================================================================
//
// ## S4E5 wire format (v0.8 #52, **read-only as of v0.8.1 #57**)
//
// ```text
// magic         4B    "S4E5"
// algo          1B    0x01 (AES-256-GCM)
// key_id        2B    BE — keyring slot the active key was at PUT time
// reserved      1B    0x00
// chunk_size    4B    BE — plaintext bytes per chunk (final chunk may be smaller)
// chunk_count   4B    BE — total chunks (always >= 1; empty plaintext = 1 zero-byte chunk)
// salt          4B    random per-PUT, mixed into every nonce
// [chunk_count] × {
//   tag         16B   AES-GCM auth tag for this chunk
//   ciphertext  N B   chunk_size bytes (final chunk: 0..=chunk_size bytes)
// }
// ```
//
// Fixed header = 20 bytes ([`S4E5_HEADER_BYTES`]).
//
// ## S4E6 wire format (v0.8.1 #57, current PUT emit)
//
// ```text
// magic         4B    "S4E6"
// algo          1B    0x01 (AES-256-GCM)
// key_id        2B    BE
// reserved      1B    0x00
// chunk_size    4B    BE
// chunk_count   4B    BE
// salt          8B    random per-PUT  ← 4B → 8B widened
// [chunk_count] × { tag 16B, ciphertext N B }
// ```
//
// Fixed header = 24 bytes ([`S4E6_HEADER_BYTES`]). Chunk array
// layout is byte-identical to S4E5; only the header (salt 4 → 8)
// and the nonce/AAD derivation differ.
//
// ## Per-chunk overhead (both S4E5 and S4E6)
//
// 16 bytes — just the AES-GCM auth tag. AES-GCM is CTR-mode, so
// `ciphertext.len() == plaintext.len()`. Total overhead for an
// N-byte plaintext at chunk size C: `header + ceil(N/C) * 16`.
//
// ## S4E5 nonce / AAD (read-only)
//
// ```text
// nonce_v5[0..4]  = b"E5\x00\x00"
// nonce_v5[4..8]  = salt (4 B)
// nonce_v5[8..12] = chunk_index BE (u32)
//
// aad_v5 = b"S4E5" || algo (1) || chunk_index BE (4) || total BE (4)
//        || key_id BE (2) || salt (4)
// ```
//
// Birthday-collision threshold on the 4-byte salt: ~50% at ~65,536
// distinct PUTs under the same key — the security regression that
// motivated #57.
//
// ## S4E6 nonce / AAD (current emit)
//
// ```text
// nonce_v6[0]     = b'E'                   (1 B fixed prefix)
// nonce_v6[1..9]  = salt (8 B)             (per-PUT random from OsRng)
// nonce_v6[9..12] = chunk_index BE (u24)   (3 B → max 16_777_215 chunks)
//
// aad_v6 = b"S4E6" || algo (1) || chunk_index BE (4) || total BE (4)
//        || key_id BE (2) || salt (8)
// ```
//
// Wider salt: birthday collision ~50% at ~2^32 = ~4.3 billion
// PUTs/key — four orders of magnitude over S4E5.
//
// chunk_index narrows from 32-bit to 24-bit, capping `chunk_count`
// at `2^24 - 1 = 16_777_215`. At the default `--sse-chunk-size
// 1048576` (1 MiB) that's ~16 PiB per object — three orders of
// magnitude over S3's 5 GiB single-object cap. Smaller chunk sizes
// need to be sized carefully: e.g. `--sse-chunk-size 64` on a
// > 1 GiB object would exceed the cap (1 GiB / 64 B = 16M+1
// chunks); such configurations surface
// [`SseError::ChunkCountTooLarge`] at PUT time rather than
// silently truncating.
//
// AAD on both variants includes the chunk index + total so chunk
// reordering or dropping fails the per-chunk tag, plus key_id +
// salt so header tampering also fails auth.

/// Fixed header size of an S4E5 frame, before any chunks. `magic 4 +
/// algo 1 + key_id 2 + reserved 1 + chunk_size 4 + chunk_count 4 +
/// salt 4` = 20 bytes.
pub const S4E5_HEADER_BYTES: usize = 4 + 1 + 2 + 1 + 4 + 4 + 4; // = 20

/// Per-chunk overhead inside an S4E5 / S4E6 frame: just the AES-GCM
/// auth tag. `ciphertext.len() == plaintext.len()` (CTR mode), so a
/// chunk of N plaintext bytes costs N + 16 on disk.
pub const S4E5_PER_CHUNK_OVERHEAD: usize = TAG_LEN; // = 16

/// v0.8.1 #57: fixed header size of an S4E6 frame. Same layout as
/// S4E5 except the per-PUT salt widens 4 → 8 bytes: `magic 4 + algo
/// 1 + key_id 2 + reserved 1 + chunk_size 4 + chunk_count 4 + salt
/// 8` = 24 bytes.
pub const S4E6_HEADER_BYTES: usize = 4 + 1 + 2 + 1 + 4 + 4 + 8; // = 24

/// v0.8.1 #57: per-chunk overhead for S4E6. Identical to S4E5
/// (same AES-GCM tag size). Re-exported as a distinct const so call
/// sites that compute on-disk size for S4E6 specifically can spell
/// the magic clearly in their arithmetic.
pub const S4E6_PER_CHUNK_OVERHEAD: usize = TAG_LEN; // = 16

/// v0.8.1 #57: maximum `chunk_count` that fits in the S4E6 nonce's
/// 24-bit chunk_index field. At 1 MiB chunks this is ~16 PiB per
/// object — three orders of magnitude over S3's 5 GiB single-object
/// cap, so it's not a practical limit at the default chunk size.
pub const S4E6_MAX_CHUNK_COUNT: u32 = (1u32 << 24) - 1; // 16_777_215

/// 4-byte fixed prefix of every S4E5 nonce. Distinct from the bytes
/// a random S4E1/E2 nonce could plausibly start with so debugging
/// dumps can immediately tell "this is a chunked nonce" from the
/// first 4 bytes.
const S4E5_NONCE_TAG: [u8; 4] = [b'E', b'5', 0, 0];

/// 1-byte fixed prefix of every S4E6 nonce. Trades 3 of S4E5's 4
/// "tag" bytes for 4 extra salt bytes (4 → 8) and 0 of the chunk
/// index bytes (24-bit instead of 32-bit). The remaining `b'E'`
/// keeps debug dumps recognizable as "chunked SSE-S4 nonce".
const S4E6_NONCE_PREFIX: u8 = b'E';

/// Variant tag for the chunked-frame helpers. Selects the nonce +
/// AAD derivation (and incidentally the salt width). The
/// chunk-array layout is byte-identical for both — only the header
/// size and the nonce/AAD derivation differ.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum ChunkedVariant {
    V5,
    V6,
}

impl ChunkedVariant {
    fn header_bytes(self) -> usize {
        match self {
            ChunkedVariant::V5 => S4E5_HEADER_BYTES,
            ChunkedVariant::V6 => S4E6_HEADER_BYTES,
        }
    }
}

/// Build the per-chunk AAD for an S4E5 chunk. Includes magic + algo
/// plus the structural chunk_index/total_chunks (so chunk reordering
/// fails auth) plus key_id + salt (so header tampering — flipping
/// key_id or salt — also fails auth).
fn aad_v5(
    chunk_index: u32,
    total_chunks: u32,
    key_id: u16,
    salt: &[u8; 4],
) -> [u8; 4 + 1 + 4 + 4 + 2 + 4] {
    let mut aad = [0u8; 4 + 1 + 4 + 4 + 2 + 4]; // = 19
    aad[..4].copy_from_slice(SSE_MAGIC_V5);
    aad[4] = ALGO_AES_256_GCM;
    aad[5..9].copy_from_slice(&chunk_index.to_be_bytes());
    aad[9..13].copy_from_slice(&total_chunks.to_be_bytes());
    aad[13..15].copy_from_slice(&key_id.to_be_bytes());
    aad[15..19].copy_from_slice(salt);
    aad
}

/// v0.8.1 #57: per-chunk AAD for S4E6. Same structural fields as
/// [`aad_v5`] (magic + algo + chunk_index + total + key_id + salt)
/// but with the wider 8-byte salt and the new `b"S4E6"` magic, so
/// an attacker can't strip the version tag and replay an S4E5
/// nonce/tag against an S4E6 frame.
fn aad_v6(
    chunk_index: u32,
    total_chunks: u32,
    key_id: u16,
    salt: &[u8; 8],
) -> [u8; 4 + 1 + 4 + 4 + 2 + 8] {
    let mut aad = [0u8; 4 + 1 + 4 + 4 + 2 + 8]; // = 23
    aad[..4].copy_from_slice(SSE_MAGIC_V6);
    aad[4] = ALGO_AES_256_GCM;
    aad[5..9].copy_from_slice(&chunk_index.to_be_bytes());
    aad[9..13].copy_from_slice(&total_chunks.to_be_bytes());
    aad[13..15].copy_from_slice(&key_id.to_be_bytes());
    aad[15..23].copy_from_slice(salt);
    aad
}

/// Derive the 12-byte AES-GCM nonce for chunk `chunk_index` from the
/// per-PUT `salt`. Pure function; no RNG state — the same `(salt,
/// chunk_index)` always yields the same nonce, which is the whole
/// point: GET reads `salt` from the header and walks the chunks
/// without storing 12 bytes of nonce per chunk.
fn nonce_v5(salt: &[u8; 4], chunk_index: u32) -> [u8; NONCE_LEN] {
    let mut n = [0u8; NONCE_LEN];
    n[..4].copy_from_slice(&S4E5_NONCE_TAG);
    n[4..8].copy_from_slice(salt);
    n[8..12].copy_from_slice(&chunk_index.to_be_bytes());
    n
}

/// v0.8.1 #57: derive the 12-byte AES-GCM nonce for an S4E6 chunk:
/// `b'E'(1) || salt(8) || chunk_index_BE_u24(3)`. The 24-bit
/// chunk_index caps `chunk_count` at 16,777,215 — see
/// [`S4E6_MAX_CHUNK_COUNT`]. The pre-encrypt path enforces this cap
/// and surfaces [`SseError::ChunkCountTooLarge`], so this function
/// only ever sees `chunk_index <= 0xFF_FFFF` (the leading byte of
/// the BE u32 is dropped).
fn nonce_v6(salt: &[u8; 8], chunk_index: u32) -> [u8; NONCE_LEN] {
    debug_assert!(
        chunk_index <= S4E6_MAX_CHUNK_COUNT,
        "S4E6 chunk_index {chunk_index} exceeds 24-bit cap (caller MUST validate)",
    );
    let mut n = [0u8; NONCE_LEN];
    n[0] = S4E6_NONCE_PREFIX;
    n[1..9].copy_from_slice(salt);
    let be = chunk_index.to_be_bytes(); // [b3, b2, b1, b0] of u32
    // Take the low 3 bytes (b2, b1, b0) — the high byte is 0 by the
    // S4E6_MAX_CHUNK_COUNT cap above.
    n[9..12].copy_from_slice(&be[1..4]);
    n
}

/// v0.8 #52 / v0.8.1 #57: encrypt `plaintext` under `keyring`'s
/// active key, sliced into independently-sealed AES-GCM chunks of
/// `chunk_size` plaintext bytes each. Returns the on-the-wire
/// **S4E6** frame (v0.8.1 #57 widened the per-PUT salt 4 B → 8 B;
/// the S4E5 emit path was retired but the [`decrypt`] /
/// [`decrypt_chunked_stream`] paths still read S4E5 objects for
/// back-compat).
///
/// Errors:
/// - [`SseError::ChunkSizeInvalid`] if `chunk_size == 0`.
/// - [`SseError::ChunkCountTooLarge`] if
///   `ceil(plaintext.len() / chunk_size) > 16_777_215` (the S4E6
///   24-bit chunk_index cap; pick a larger `--sse-chunk-size`).
///
/// Empty plaintext is permitted and produces a frame with
/// `chunk_count = 1, ciphertext_len = 0` (one all-tag chunk). That
/// keeps the GET chunk-walk loop simpler — it never has to
/// special-case zero chunks.
///
/// `chunk_size` is the *plaintext* bytes per chunk; the on-disk
/// ciphertext per chunk is the same number (AES-GCM is CTR-mode),
/// plus the 16-byte tag prepended.
pub fn encrypt_v2_chunked(
    plaintext: &[u8],
    keyring: &SseKeyring,
    chunk_size: usize,
) -> Result<Bytes, SseError> {
    if chunk_size == 0 {
        return Err(SseError::ChunkSizeInvalid);
    }
    let (key_id, key) = keyring.active();
    let cipher = Aes256Gcm::new(key.as_aes_key());
    let mut salt = [0u8; 8];
    rand::rngs::OsRng.fill_bytes(&mut salt);

    // Always emit at least one chunk (so an empty plaintext still
    // has a well-defined header → chunk_count >= 1 invariant).
    let chunk_count_usize = if plaintext.is_empty() {
        1
    } else {
        plaintext.len().div_ceil(chunk_size)
    };
    // Saturating-cast to u32 so we report ChunkCountTooLarge cleanly
    // for inputs that would overflow u32 too (would need a > 16 EiB
    // plaintext at chunk_size = 1 — astronomical, but defensive).
    let chunk_count: u32 = u32::try_from(chunk_count_usize).unwrap_or(u32::MAX);
    if chunk_count > S4E6_MAX_CHUNK_COUNT {
        return Err(SseError::ChunkCountTooLarge {
            got: chunk_count,
            max: S4E6_MAX_CHUNK_COUNT,
        });
    }

    let mut out = Vec::with_capacity(
        S4E6_HEADER_BYTES + plaintext.len() + (chunk_count as usize * S4E6_PER_CHUNK_OVERHEAD),
    );
    out.extend_from_slice(SSE_MAGIC_V6);
    out.push(ALGO_AES_256_GCM);
    out.extend_from_slice(&key_id.to_be_bytes());
    out.push(0u8); // reserved
    out.extend_from_slice(&(chunk_size as u32).to_be_bytes());
    out.extend_from_slice(&chunk_count.to_be_bytes());
    out.extend_from_slice(&salt);

    for i in 0..chunk_count {
        let off = (i as usize).saturating_mul(chunk_size);
        let end = off.saturating_add(chunk_size).min(plaintext.len());
        let chunk_pt: &[u8] = if off >= plaintext.len() {
            // Empty-plaintext / past-end (only the single-chunk
            // empty-plaintext case lands here).
            &[]
        } else {
            &plaintext[off..end]
        };
        let nonce_bytes = nonce_v6(&salt, i);
        let nonce = Nonce::from_slice(&nonce_bytes);
        let aad = aad_v6(i, chunk_count, key_id, &salt);
        let ct_with_tag = cipher
            .encrypt(
                nonce,
                Payload {
                    msg: chunk_pt,
                    aad: &aad,
                },
            )
            .expect("aes-gcm encrypt cannot fail with a 32-byte key");
        debug_assert!(ct_with_tag.len() >= TAG_LEN);
        let split = ct_with_tag.len() - TAG_LEN;
        let (ct, tag) = ct_with_tag.split_at(split);
        out.extend_from_slice(tag);
        out.extend_from_slice(ct);
        crate::metrics::record_sse_streaming_chunk("encrypt");
    }
    Ok(Bytes::from(out))
}

/// Salt material for a chunked frame — branches on variant so the
/// shared chunk-walking loop can carry both 4-byte (S4E5) and
/// 8-byte (S4E6) salts without an extra heap alloc.
#[derive(Debug, Clone, Copy)]
enum ChunkedSalt {
    V5([u8; 4]),
    V6([u8; 8]),
}

/// Parsed S4E5 / S4E6 header — fixed-layout fields. Used by the
/// buffered ([`decrypt_chunked_buffered`]) and streaming
/// ([`decrypt_chunked_stream`]) paths to share frame validation
/// across both variants.
#[derive(Debug, Clone, Copy)]
struct ChunkedHeader {
    /// Used only by tests today (asserts on which frame variant
    /// parsed); in production the variant is implicit in
    /// `salt`'s ChunkedSalt arm. Kept as a field rather than
    /// re-deriving from `salt` so the parser writes one source of
    /// truth.
    #[allow(dead_code)]
    variant: ChunkedVariant,
    key_id: u16,
    chunk_size: u32,
    chunk_count: u32,
    salt: ChunkedSalt,
    /// Byte offset where the chunk array starts (always
    /// `variant.header_bytes()`; carried in the struct so call sites
    /// don't have to re-derive it from the variant tag).
    chunks_offset: usize,
}

/// Parsed view of an S4E6 frame's fixed header. Public mirror of
/// the S4E4 parser — useful for admin tools or future inspectors
/// that want to enumerate object → key_id bindings without
/// re-implementing the offset math. The `salt` borrow keeps
/// allocations to zero (the slice points back into the input
/// buffer).
#[derive(Debug, Clone, Copy)]
pub struct S4E6Header<'a> {
    pub key_id: u16,
    pub chunk_size: u32,
    pub chunk_count: u32,
    pub salt: &'a [u8; 8],
}

/// Pure byte-shuffle parser for an S4E6 fixed header (24 bytes). No
/// crypto, no keyring lookup. Errors on truncation, wrong magic,
/// unsupported algo, or zero `chunk_size` / `chunk_count`.
pub fn parse_s4e6_header(blob: &[u8]) -> Result<S4E6Header<'_>, SseError> {
    if blob.len() < S4E6_HEADER_BYTES {
        return Err(SseError::ChunkFrameTruncated { what: "header" });
    }
    if &blob[..4] != SSE_MAGIC_V6 {
        let mut got = [0u8; 4];
        got.copy_from_slice(&blob[..4]);
        return Err(SseError::BadMagic { got });
    }
    let algo = blob[4];
    if algo != ALGO_AES_256_GCM {
        return Err(SseError::UnsupportedAlgo { tag: algo });
    }
    let key_id = u16::from_be_bytes([blob[5], blob[6]]);
    // blob[7] = reserved (0; authenticated as 0 via AAD).
    let chunk_size = u32::from_be_bytes([blob[8], blob[9], blob[10], blob[11]]);
    let chunk_count = u32::from_be_bytes([blob[12], blob[13], blob[14], blob[15]]);
    if chunk_size == 0 {
        return Err(SseError::ChunkSizeInvalid);
    }
    if chunk_count == 0 {
        return Err(SseError::ChunkFrameTruncated {
            what: "chunk_count == 0",
        });
    }
    if chunk_count > S4E6_MAX_CHUNK_COUNT {
        return Err(SseError::ChunkCountTooLarge {
            got: chunk_count,
            max: S4E6_MAX_CHUNK_COUNT,
        });
    }
    let salt: &[u8; 8] = (&blob[16..24]).try_into().expect("8B salt slice");
    Ok(S4E6Header {
        key_id,
        chunk_size,
        chunk_count,
        salt,
    })
}

fn parse_chunked_header(body: &[u8], max_body_bytes: usize) -> Result<ChunkedHeader, SseError> {
    if body.len() < 4 {
        return Err(SseError::ChunkFrameTruncated { what: "magic" });
    }
    let magic = &body[..4];
    let variant = if magic == SSE_MAGIC_V5 {
        ChunkedVariant::V5
    } else if magic == SSE_MAGIC_V6 {
        ChunkedVariant::V6
    } else {
        let mut got = [0u8; 4];
        got.copy_from_slice(magic);
        return Err(SseError::BadMagic { got });
    };
    let header_bytes = variant.header_bytes();
    if body.len() < header_bytes {
        return Err(SseError::ChunkFrameTruncated { what: "header" });
    }
    let algo = body[4];
    if algo != ALGO_AES_256_GCM {
        return Err(SseError::UnsupportedAlgo { tag: algo });
    }
    let key_id = u16::from_be_bytes([body[5], body[6]]);
    // body[7] = reserved (must be 0; authenticated as 0 via AAD).
    let chunk_size = u32::from_be_bytes([body[8], body[9], body[10], body[11]]);
    let chunk_count = u32::from_be_bytes([body[12], body[13], body[14], body[15]]);
    if chunk_size == 0 {
        return Err(SseError::ChunkSizeInvalid);
    }
    if chunk_count == 0 {
        return Err(SseError::ChunkFrameTruncated {
            what: "chunk_count == 0",
        });
    }
    let salt = match variant {
        ChunkedVariant::V5 => {
            let mut s = [0u8; 4];
            s.copy_from_slice(&body[16..20]);
            ChunkedSalt::V5(s)
        }
        ChunkedVariant::V6 => {
            // v0.8.1 #57 sanity check: the encoder enforces this cap,
            // but a tampered / malicious frame could declare a huge
            // chunk_count that would loop the walker 16M+ times if
            // we trusted it. Reject early.
            if chunk_count > S4E6_MAX_CHUNK_COUNT {
                return Err(SseError::ChunkCountTooLarge {
                    got: chunk_count,
                    max: S4E6_MAX_CHUNK_COUNT,
                });
            }
            let mut s = [0u8; 8];
            s.copy_from_slice(&body[16..24]);
            ChunkedSalt::V6(s)
        }
    };

    // v0.8.2 #64 — DoS guard. Pre-validate the declared
    // (chunk_size × chunk_count) plaintext size in u64 before *any*
    // downstream allocation or chunk-walk loop. Rejects:
    //   1. arithmetic that would overflow u64 (e.g. chunk_size =
    //      u32::MAX, chunk_count = u32::MAX → 2^64 ≫ u64::MAX),
    //   2. a body that is *larger* than the maximum a header with
    //      these declared sizes could plausibly carry (i.e. trailing
    //      bytes after the final chunk → corruption / append attack),
    //      and
    //   3. a plaintext size larger than the caller's configured
    //      `max_body_bytes` (default 5 GiB).
    //
    // We do NOT require the body to be `>= max_total` here — the
    // final chunk may legitimately hold fewer than `chunk_size`
    // plaintext bytes (encoder uses
    // `chunk_count = ceil(plaintext.len() / chunk_size)`, so the last
    // chunk holds the remainder). The walker handles that
    // variable-length tail; truncation of the final chunk's tag /
    // ciphertext surfaces as `ChunkFrameTruncated` once the walker
    // gets there. The pre-alloc guard's job is the *upper-bound*
    // checks: kill obviously-impossible shapes before we
    // `Vec::with_capacity(chunk_size × chunk_count)`.
    //
    // All paths return concrete error variants; nothing panics on
    // adversarial input. The error strings carry only constant
    // operator-readable labels (no echoed numbers) so the response is
    // not a tampering oracle for the attacker.
    let chunk_size_u64 = chunk_size as u64;
    let chunk_count_u64 = chunk_count as u64;
    let expected_plain_size =
        chunk_size_u64
            .checked_mul(chunk_count_u64)
            .ok_or(SseError::ChunkFrameTooLarge {
                details: "chunk_size * chunk_count overflows u64",
            })?;
    let per_chunk_overhead = S4E5_PER_CHUNK_OVERHEAD as u64; // = 16 (AES-GCM tag)
    let total_tag_overhead =
        per_chunk_overhead
            .checked_mul(chunk_count_u64)
            .ok_or(SseError::ChunkFrameTooLarge {
                details: "tag_len * chunk_count overflows u64",
            })?;
    let max_total = expected_plain_size
        .checked_add(total_tag_overhead)
        .and_then(|t| t.checked_add(header_bytes as u64))
        .ok_or(SseError::ChunkFrameTooLarge {
            details: "header + plaintext + tag overhead overflows u64",
        })?;
    // Body cannot be *larger* than what the declared geometry could
    // legitimately produce. Any extra bytes past the final chunk are
    // either trailing junk (corruption) or an append attack — kill
    // them here so the walker doesn't even start verifying chunks
    // for an obviously-malformed body. (The walker's own end-of-loop
    // `cursor != body.len()` check also catches this; the pre-alloc
    // guard saves the chunk-walk worth of AES-GCM verifies in the
    // hostile case.)
    if (body.len() as u64) > max_total {
        return Err(SseError::ChunkFrameTruncated {
            what: "trailing bytes past declared chunk geometry",
        });
    }
    // The actual DoS guard: cap on declared plaintext. If the header
    // claims more than the operator's configured single-object
    // ceiling, refuse before the buffered path's
    // `Vec::with_capacity(chunk_size * chunk_count)` runs.
    if expected_plain_size > max_body_bytes as u64 {
        return Err(SseError::ChunkFrameTooLarge {
            details: "declared plaintext exceeds gateway max_body_bytes",
        });
    }

    Ok(ChunkedHeader {
        variant,
        key_id,
        chunk_size,
        chunk_count,
        salt,
        chunks_offset: header_bytes,
    })
}

/// Decrypt one chunk under either the S4E5 or S4E6 derivation. Used
/// by both the buffered and streaming paths so AAD / nonce
/// derivation lives in exactly one place.
fn decrypt_chunked_chunk(
    cipher: &Aes256Gcm,
    chunk_index: u32,
    chunk_count: u32,
    key_id: u16,
    salt: &ChunkedSalt,
    tag: &[u8; TAG_LEN],
    ct: &[u8],
) -> Result<Bytes, SseError> {
    let nonce_bytes = match salt {
        ChunkedSalt::V5(s) => nonce_v5(s, chunk_index),
        ChunkedSalt::V6(s) => nonce_v6(s, chunk_index),
    };
    let nonce = Nonce::from_slice(&nonce_bytes);
    let mut ct_with_tag = Vec::with_capacity(ct.len() + TAG_LEN);
    ct_with_tag.extend_from_slice(ct);
    ct_with_tag.extend_from_slice(tag);
    let result = match salt {
        ChunkedSalt::V5(s) => {
            let aad = aad_v5(chunk_index, chunk_count, key_id, s);
            cipher.decrypt(
                nonce,
                Payload {
                    msg: &ct_with_tag,
                    aad: &aad,
                },
            )
        }
        ChunkedSalt::V6(s) => {
            let aad = aad_v6(chunk_index, chunk_count, key_id, s);
            cipher.decrypt(
                nonce,
                Payload {
                    msg: &ct_with_tag,
                    aad: &aad,
                },
            )
        }
    };
    result
        .map(Bytes::from)
        .map_err(|_| SseError::ChunkAuthFailed { chunk_index })
}

/// Walk an S4E5 / S4E6 body chunk-by-chunk, calling `emit` on each
/// successfully-verified plaintext chunk. Returns immediately on the
/// first chunk that fails auth or is truncated. Shared core between
/// the buffered ([`decrypt_chunked_buffered`]) and streaming
/// ([`decrypt_chunked_stream`]) paths.
fn walk_chunked<F: FnMut(Bytes) -> Result<(), SseError>>(
    body: &[u8],
    keyring: &SseKeyring,
    max_body_bytes: usize,
    mut emit: F,
) -> Result<(), SseError> {
    let hdr = parse_chunked_header(body, max_body_bytes)?;
    let key = keyring
        .get(hdr.key_id)
        .ok_or(SseError::KeyNotInKeyring { id: hdr.key_id })?;
    let cipher = Aes256Gcm::new(key.as_aes_key());

    let mut cursor = hdr.chunks_offset;
    let chunk_size = hdr.chunk_size as usize;
    for i in 0..hdr.chunk_count {
        if cursor + TAG_LEN > body.len() {
            return Err(SseError::ChunkFrameTruncated { what: "chunk tag" });
        }
        let tag_off = cursor;
        let ct_off = tag_off + TAG_LEN;
        let is_last = i + 1 == hdr.chunk_count;
        let ct_len = if is_last {
            if ct_off > body.len() {
                return Err(SseError::ChunkFrameTruncated {
                    what: "final chunk ciphertext",
                });
            }
            let remaining = body.len() - ct_off;
            if remaining > chunk_size {
                return Err(SseError::ChunkFrameTruncated {
                    what: "trailing bytes after final chunk",
                });
            }
            remaining
        } else {
            chunk_size
        };
        let ct_end = ct_off + ct_len;
        if ct_end > body.len() {
            return Err(SseError::ChunkFrameTruncated {
                what: "chunk ciphertext",
            });
        }
        let mut tag = [0u8; TAG_LEN];
        tag.copy_from_slice(&body[tag_off..ct_off]);
        let ct = &body[ct_off..ct_end];
        let plain =
            decrypt_chunked_chunk(&cipher, i, hdr.chunk_count, hdr.key_id, &hdr.salt, &tag, ct)?;
        crate::metrics::record_sse_streaming_chunk("decrypt");
        emit(plain)?;
        cursor = ct_end;
    }
    if cursor != body.len() {
        return Err(SseError::ChunkFrameTruncated {
            what: "trailing bytes after declared chunk_count",
        });
    }
    Ok(())
}

/// Sync back-compat path: decrypt every chunk and concatenate into
/// a single `Bytes`. Memory peak = full plaintext (defeats the
/// point of S4E5/S4E6 streaming, but useful for callers that already
/// need the whole body — e.g. server-side restream-rewrite paths or
/// unit tests). Accepts both S4E5 (legacy) and S4E6 (current) bodies.
///
/// `max_body_bytes` (v0.8.2 #64) caps the declared plaintext size in
/// the header **before any allocation**. A malicious / corrupted
/// frame that claims a multi-PB plaintext is rejected as
/// [`SseError::ChunkFrameTooLarge`] with no `Vec::with_capacity` call
/// behind it. Callers without a bespoke cap should use
/// [`decrypt_chunked_buffered_default`] (5 GiB cap).
pub fn decrypt_chunked_buffered(
    body: &[u8],
    keyring: &SseKeyring,
    max_body_bytes: usize,
) -> Result<Bytes, SseError> {
    let hdr = parse_chunked_header(body, max_body_bytes)?;
    // Header is validated above: chunk_size * chunk_count fits u64
    // and is ≤ max_body_bytes (≤ 5 GiB on this gateway), so the
    // `as usize` casts cannot truncate on a 64-bit host. The
    // `Vec::with_capacity` is therefore bounded by the same cap that
    // protects the rest of the gateway from oversized PUTs / GETs.
    let mut out = Vec::with_capacity(hdr.chunk_size as usize * hdr.chunk_count as usize);
    walk_chunked(body, keyring, max_body_bytes, |chunk| {
        out.extend_from_slice(&chunk);
        Ok(())
    })?;
    Ok(Bytes::from(out))
}

/// Back-compat wrapper around [`decrypt_chunked_buffered`] that uses
/// [`DEFAULT_MAX_BODY_BYTES`] (5 GiB — AWS S3's single-object PUT
/// ceiling). Provided so the existing `decrypt(body, keyring)`
/// dispatcher and call sites that don't have a bespoke cap can keep
/// their signature unchanged after the v0.8.2 #64 pre-allocation
/// guard landed.
pub fn decrypt_chunked_buffered_default(
    body: &[u8],
    keyring: &SseKeyring,
) -> Result<Bytes, SseError> {
    decrypt_chunked_buffered(body, keyring, DEFAULT_MAX_BODY_BYTES)
}

/// v0.8 #52 (S4E5) / v0.8.1 #57 (S4E6): stream-decrypt API for
/// chunked SSE-S4 bodies. Returns a [`futures::Stream`] that yields
/// one `Bytes` per chunk in order. Each chunk is emitted only after
/// AES-GCM tag verify succeeds, so the client never sees plaintext
/// bytes that haven't been authenticated. A failing chunk yields
/// its [`SseError::ChunkAuthFailed`] (with the chunk index) and ends
/// the stream — earlier chunks may already have left the gateway,
/// which matches the standard streaming-AEAD trade-off (operators
/// MUST alert on the audit log + metric, not rely on connection
/// close to guarantee atomicity).
///
/// Accepts either S4E5 (v0.8 #52, legacy) or S4E6 (v0.8.1 #57,
/// current) magic. Non-chunked magic surfaces as
/// [`SseError::BadMagic`] / [`SseError::ChunkFrameTruncated`] on
/// the first poll — the stream is "fail-fast" rather than "fall
/// through to S4E2 buffered decrypt", because the caller has
/// already dispatched on [`peek_magic`] by the time it hands a body
/// to this function.
///
/// `body` is owned by the returned stream so the caller doesn't
/// need to keep the bytes alive separately. The returned stream is
/// `'static` — the `keyring` borrow is consumed up front to extract
/// the per-frame key and build the AES cipher (which owns its key
/// material), so the caller's keyring may be dropped immediately.
pub fn decrypt_chunked_stream(
    body: bytes::Bytes,
    keyring: &SseKeyring,
) -> impl futures::Stream<Item = Result<Bytes, SseError>> + 'static {
    use futures::stream::{self, StreamExt};

    // Cheap pre-validation: parse the header + look up the key
    // once, up front, so a malformed frame surfaces on the first
    // poll instead of being deferred behind the first-chunk loop.
    // The `keyring` borrow ends here — we extract the AES key into
    // the owned `Aes256Gcm` cipher, then store that in the stream
    // state.
    let prelude = (|| {
        // v0.8.2 #64: streaming path is naturally memory-bounded
        // (one chunk-worth of plaintext at a time), so we don't cap
        // the *declared* total here — the per-chunk loop already
        // bounds memory by `chunk_size`. The truncation /
        // arithmetic-overflow checks inside `parse_chunked_header`
        // still run regardless of `max_body_bytes`, which is what
        // protects the streaming path from the same DoS class as the
        // buffered path (the overflow / declared-vs-actual length
        // mismatches are independent of the gateway cap).
        let hdr = parse_chunked_header(&body, usize::MAX)?;
        let key = keyring
            .get(hdr.key_id)
            .ok_or(SseError::KeyNotInKeyring { id: hdr.key_id })?;
        let cipher = Aes256Gcm::new(key.as_aes_key());
        Ok::<_, SseError>((hdr, cipher))
    })();

    match prelude {
        Err(e) => stream::iter(std::iter::once(Err(e))).left_stream(),
        Ok((hdr, cipher)) => {
            let chunks_offset = hdr.chunks_offset;
            let state = ChunkedDecryptState {
                body,
                cipher,
                hdr,
                cursor: chunks_offset,
                next_index: 0,
            };
            stream::try_unfold(state, decrypt_next_chunk).right_stream()
        }
    }
}

/// Per-stream state for [`decrypt_chunked_stream`]. Holds the owned
/// `body` (so the stream stays self-contained), the prepared
/// cipher, and the cursor position into the chunk array.
struct ChunkedDecryptState {
    body: bytes::Bytes,
    cipher: Aes256Gcm,
    hdr: ChunkedHeader,
    cursor: usize,
    next_index: u32,
}

async fn decrypt_next_chunk(
    mut state: ChunkedDecryptState,
) -> Result<Option<(Bytes, ChunkedDecryptState)>, SseError> {
    if state.next_index >= state.hdr.chunk_count {
        // Final boundary check — anything past the declared
        // chunk_count would be a truncation / append attack.
        if state.cursor != state.body.len() {
            return Err(SseError::ChunkFrameTruncated {
                what: "trailing bytes after declared chunk_count",
            });
        }
        return Ok(None);
    }
    let i = state.next_index;
    let chunk_size = state.hdr.chunk_size as usize;
    if state.cursor + TAG_LEN > state.body.len() {
        return Err(SseError::ChunkFrameTruncated { what: "chunk tag" });
    }
    let tag_off = state.cursor;
    let ct_off = tag_off + TAG_LEN;
    let is_last = i + 1 == state.hdr.chunk_count;
    let ct_len = if is_last {
        if ct_off > state.body.len() {
            return Err(SseError::ChunkFrameTruncated {
                what: "final chunk ciphertext",
            });
        }
        let remaining = state.body.len() - ct_off;
        if remaining > chunk_size {
            return Err(SseError::ChunkFrameTruncated {
                what: "trailing bytes after final chunk",
            });
        }
        remaining
    } else {
        chunk_size
    };
    let ct_end = ct_off + ct_len;
    if ct_end > state.body.len() {
        return Err(SseError::ChunkFrameTruncated {
            what: "chunk ciphertext",
        });
    }
    let mut tag = [0u8; TAG_LEN];
    tag.copy_from_slice(&state.body[tag_off..ct_off]);
    let ct = &state.body[ct_off..ct_end];
    let plain = decrypt_chunked_chunk(
        &state.cipher,
        i,
        state.hdr.chunk_count,
        state.hdr.key_id,
        &state.hdr.salt,
        &tag,
        ct,
    )?;
    crate::metrics::record_sse_streaming_chunk("decrypt");
    state.cursor = ct_end;
    state.next_index += 1;
    Ok(Some((plain, state)))
}

/// v0.8.1 #57: build an S4E5 frame. Identical body structure to the
/// pre-#57 `encrypt_v2_chunked` (4-byte salt + S4E5 magic + V5
/// nonce/AAD); kept around purely so the back-compat-read tests can
/// synthesize a "v0.8.0 vintage" blob and prove the new gateway
/// still decrypts it.
#[cfg(test)]
fn encrypt_v2_chunked_s4e5_for_test(
    plaintext: &[u8],
    keyring: &SseKeyring,
    chunk_size: usize,
) -> Result<Bytes, SseError> {
    if chunk_size == 0 {
        return Err(SseError::ChunkSizeInvalid);
    }
    let (key_id, key) = keyring.active();
    let cipher = Aes256Gcm::new(key.as_aes_key());
    let mut salt = [0u8; 4];
    rand::rngs::OsRng.fill_bytes(&mut salt);

    let chunk_count: u32 = if plaintext.is_empty() {
        1
    } else {
        plaintext
            .len()
            .div_ceil(chunk_size)
            .try_into()
            .expect("chunk_count overflows u32")
    };

    let mut out = Vec::with_capacity(
        S4E5_HEADER_BYTES + plaintext.len() + (chunk_count as usize * S4E5_PER_CHUNK_OVERHEAD),
    );
    out.extend_from_slice(SSE_MAGIC_V5);
    out.push(ALGO_AES_256_GCM);
    out.extend_from_slice(&key_id.to_be_bytes());
    out.push(0u8);
    out.extend_from_slice(&(chunk_size as u32).to_be_bytes());
    out.extend_from_slice(&chunk_count.to_be_bytes());
    out.extend_from_slice(&salt);

    for i in 0..chunk_count {
        let off = (i as usize).saturating_mul(chunk_size);
        let end = off.saturating_add(chunk_size).min(plaintext.len());
        let chunk_pt: &[u8] = if off >= plaintext.len() {
            &[]
        } else {
            &plaintext[off..end]
        };
        let nonce_bytes = nonce_v5(&salt, i);
        let nonce = Nonce::from_slice(&nonce_bytes);
        let aad = aad_v5(i, chunk_count, key_id, &salt);
        let ct_with_tag = cipher
            .encrypt(
                nonce,
                Payload {
                    msg: chunk_pt,
                    aad: &aad,
                },
            )
            .expect("aes-gcm encrypt cannot fail with a 32-byte key");
        let split = ct_with_tag.len() - TAG_LEN;
        let (ct, tag) = ct_with_tag.split_at(split);
        out.extend_from_slice(tag);
        out.extend_from_slice(ct);
    }
    Ok(Bytes::from(out))
}

#[cfg(test)]
mod tests {
    use super::*;

    fn key32(seed: u8) -> Arc<SseKey> {
        Arc::new(SseKey::from_bytes(&[seed; 32]).unwrap())
    }

    fn keyring_single(seed: u8) -> SseKeyring {
        SseKeyring::new(1, key32(seed))
    }

    #[test]
    fn roundtrip_basic_v1() {
        // back-compat single-key API — still works.
        let k = SseKey::from_bytes(&[7u8; 32]).unwrap();
        let pt = b"the quick brown fox jumps over the lazy dog";
        let ct = encrypt(&k, pt);
        assert!(looks_encrypted(&ct));
        assert_eq!(&ct[..4], SSE_MAGIC_V1);
        assert_eq!(ct[4], ALGO_AES_256_GCM);
        assert_eq!(ct.len(), SSE_HEADER_BYTES + pt.len());
        // decrypt via single-key keyring
        let kr = SseKeyring::new(1, Arc::new(k));
        let pt2 = decrypt(&ct, &kr).unwrap();
        assert_eq!(pt2.as_ref(), pt);
    }

    #[test]
    fn s4e2_roundtrip_active_key() {
        let kr = keyring_single(7);
        let pt = b"S4E2 active-key roundtrip";
        let ct = encrypt_v2(pt, &kr);
        assert_eq!(&ct[..4], SSE_MAGIC_V2);
        assert_eq!(ct[4], ALGO_AES_256_GCM);
        assert_eq!(u16::from_be_bytes([ct[5], ct[6]]), 1, "key_id BE");
        assert_eq!(ct[7], 0, "reserved byte");
        assert_eq!(ct.len(), SSE_HEADER_BYTES + pt.len());
        assert!(looks_encrypted(&ct));
        let pt2 = decrypt(&ct, &kr).unwrap();
        assert_eq!(pt2.as_ref(), pt);
    }

    #[test]
    fn decrypt_s4e1_via_active_only_keyring() {
        // v0.4 wrote S4E1 with key K; v0.5 keyring has K as the only
        // (active) key. Decrypt must succeed.
        let k_arc = key32(11);
        let legacy_ct = encrypt(&k_arc, b"v0.4 vintage object");
        assert_eq!(&legacy_ct[..4], SSE_MAGIC_V1);
        let kr = SseKeyring::new(1, Arc::clone(&k_arc));
        let plain = decrypt(&legacy_ct, &kr).unwrap();
        assert_eq!(plain.as_ref(), b"v0.4 vintage object");
    }

    #[test]
    fn decrypt_s4e2_under_old_key_after_rotation() {
        // Rotation flow: object was encrypted under key id=1 when 1
        // was active. Operator rotates to active=2 and keeps 1 in the
        // keyring. The S4E2 body must still decrypt.
        let k1 = key32(1);
        let k2 = key32(2);
        let mut kr_old = SseKeyring::new(1, Arc::clone(&k1));
        let ct = encrypt_v2(b"old-rotation object", &kr_old);
        assert_eq!(u16::from_be_bytes([ct[5], ct[6]]), 1);

        // After rotation: active=2, but key 1 still in ring.
        kr_old.add(2, Arc::clone(&k2));
        let mut kr_new = SseKeyring::new(2, Arc::clone(&k2));
        kr_new.add(1, Arc::clone(&k1));

        let plain = decrypt(&ct, &kr_new).unwrap();
        assert_eq!(plain.as_ref(), b"old-rotation object");

        // And new PUTs go to id 2 (active).
        let new_ct = encrypt_v2(b"new-rotation object", &kr_new);
        assert_eq!(u16::from_be_bytes([new_ct[5], new_ct[6]]), 2);
        let plain_new = decrypt(&new_ct, &kr_new).unwrap();
        assert_eq!(plain_new.as_ref(), b"new-rotation object");
    }

    #[test]
    fn s4e2_unknown_key_id_errors() {
        let kr = keyring_single(3); // only id=1 present
        let kr_other = SseKeyring::new(99, key32(3));
        let ct = encrypt_v2(b"x", &kr_other); // body claims key_id=99
        let err = decrypt(&ct, &kr).unwrap_err();
        assert!(
            matches!(err, SseError::KeyNotInKeyring { id: 99 }),
            "got {err:?}"
        );
    }

    #[test]
    fn s4e2_tampered_key_id_fails_auth() {
        let kr = SseKeyring::new(1, key32(4));
        let mut kr_with_2 = kr.clone();
        kr_with_2.add(2, key32(5)); // a real but wrong key under id=2
        let mut ct = encrypt_v2(b"do not flip my key id", &kr).to_vec();
        // Flip key_id from 1 → 2 in the header. The keyring HAS a key
        // for 2, so the lookup succeeds — but AAD authenticates the
        // original key_id, so AES-GCM tag verification must fail.
        assert_eq!(u16::from_be_bytes([ct[5], ct[6]]), 1);
        ct[5] = 0;
        ct[6] = 2;
        let err = decrypt(&ct, &kr_with_2).unwrap_err();
        assert!(matches!(err, SseError::DecryptFailed), "got {err:?}");
    }

    #[test]
    fn s4e2_tampered_ciphertext_fails() {
        let kr = SseKeyring::new(7, key32(9));
        let mut ct = encrypt_v2(b"secret message v2", &kr).to_vec();
        let last = ct.len() - 1;
        ct[last] ^= 0x01;
        let err = decrypt(&ct, &kr).unwrap_err();
        assert!(matches!(err, SseError::DecryptFailed));
    }

    #[test]
    fn s4e2_tampered_algo_byte_fails() {
        let kr = SseKeyring::new(1, key32(2));
        let mut ct = encrypt_v2(b"hi", &kr).to_vec();
        ct[4] = 99;
        let err = decrypt(&ct, &kr).unwrap_err();
        assert!(matches!(err, SseError::UnsupportedAlgo { tag: 99 }));
    }

    #[test]
    fn wrong_key_fails_v1_via_keyring() {
        // S4E1 written under key K1; keyring has only K2 → DecryptFailed.
        let k1 = SseKey::from_bytes(&[1u8; 32]).unwrap();
        let ct = encrypt(&k1, b"secret");
        let kr_wrong = SseKeyring::new(1, Arc::new(SseKey::from_bytes(&[2u8; 32]).unwrap()));
        let err = decrypt(&ct, &kr_wrong).unwrap_err();
        assert!(matches!(err, SseError::DecryptFailed));
    }

    #[test]
    fn rejects_short_body() {
        let kr = SseKeyring::new(1, key32(1));
        let err = decrypt(b"short", &kr).unwrap_err();
        assert!(matches!(err, SseError::TooShort { got: 5 }));
    }

    #[test]
    fn looks_encrypted_passthrough_returns_false() {
        // S4F2 frame magic, NOT S4E1 / S4E2 — must not be confused.
        let f2 = b"S4F2\x01\x00\x00\x00........................................";
        assert!(!looks_encrypted(f2));
        assert!(!looks_encrypted(b""));
    }

    #[test]
    fn looks_encrypted_detects_both_v1_and_v2() {
        let kr = SseKeyring::new(1, key32(8));
        let v1 = encrypt(&SseKey::from_bytes(&[8u8; 32]).unwrap(), b"x");
        let v2 = encrypt_v2(b"x", &kr);
        assert!(looks_encrypted(&v1));
        assert!(looks_encrypted(&v2));
    }

    #[test]
    fn key_from_hex_string() {
        let bad =
            SseKey::from_bytes(b"0102030405060708090a0b0c0d0e0f10111213141516171819202122232425")
                .unwrap_err();
        assert!(matches!(bad, SseError::BadKeyLength { .. }));
        let good = b"0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef";
        let _ = SseKey::from_bytes(good).expect("64-char hex should parse");
    }

    #[test]
    fn encrypt_v2_uses_random_nonce() {
        let kr = SseKeyring::new(1, key32(3));
        let pt = b"deterministic input";
        let a = encrypt_v2(pt, &kr);
        let b = encrypt_v2(pt, &kr);
        assert_ne!(a, b, "nonce must be random per-call");
    }

    #[test]
    fn keyring_active_and_get() {
        let k1 = key32(1);
        let k2 = key32(2);
        let mut kr = SseKeyring::new(1, Arc::clone(&k1));
        kr.add(2, Arc::clone(&k2));
        let (id, active) = kr.active();
        assert_eq!(id, 1);
        assert_eq!(active.bytes, [1u8; 32]);
        assert!(kr.get(2).is_some());
        assert!(kr.get(3).is_none());
    }

    // -----------------------------------------------------------------
    // v0.5 #27 — SSE-C (customer-provided key, S4E3 frame) tests
    // -----------------------------------------------------------------

    use base64::Engine as _;

    fn cust_key(seed: u8) -> CustomerKeyMaterial {
        let key = [seed; KEY_LEN];
        let key_md5 = compute_key_md5(&key);
        CustomerKeyMaterial { key, key_md5 }
    }

    #[test]
    fn s4e3_roundtrip_happy_path() {
        let m = cust_key(42);
        let pt = b"top-secret SSE-C payload";
        let ct = encrypt_with_source(
            pt,
            SseSource::CustomerKey {
                key: &m.key,
                key_md5: &m.key_md5,
            },
        );
        // Frame inspection.
        assert_eq!(&ct[..4], SSE_MAGIC_V3);
        assert_eq!(ct[4], ALGO_AES_256_GCM);
        assert_eq!(&ct[5..5 + KEY_MD5_LEN], &m.key_md5);
        assert_eq!(ct.len(), SSE_HEADER_BYTES_V3 + pt.len());
        assert!(looks_encrypted(&ct));
        // Decrypt round-trip.
        let plain = decrypt(
            &ct,
            SseSource::CustomerKey {
                key: &m.key,
                key_md5: &m.key_md5,
            },
        )
        .unwrap();
        assert_eq!(plain.as_ref(), pt);
        // And via the From impl on &CustomerKeyMaterial.
        let plain2 = decrypt(&ct, &m).unwrap();
        assert_eq!(plain2.as_ref(), pt);
    }

    #[test]
    fn s4e3_wrong_key_yields_wrong_customer_key_error() {
        let m = cust_key(1);
        let other = cust_key(2);
        let ct = encrypt_with_source(b"payload", (&m).into());
        let err = decrypt(
            &ct,
            SseSource::CustomerKey {
                key: &other.key,
                key_md5: &other.key_md5,
            },
        )
        .unwrap_err();
        assert!(matches!(err, SseError::WrongCustomerKey), "got {err:?}");
    }

    #[test]
    fn s4e3_tampered_stored_md5_is_caught() {
        // Attacker rewrites the stored MD5 to match a key they know.
        // Even though the supplied (attacker) key matches the rewritten
        // MD5, AES-GCM authenticates the ORIGINAL md5 via AAD, so the
        // tag check fails. Surface: WrongCustomerKey if the supplied
        // md5 != stored md5 (this test), or DecryptFailed if attacker
        // also rewrites their supplied md5 to match.
        let m = cust_key(7);
        let mut ct = encrypt_with_source(b"victim payload", (&m).into()).to_vec();
        // Flip a byte in the stored fingerprint.
        ct[5] ^= 0x55;
        // Client supplies the original (unmodified) key + md5.
        let err = decrypt(
            &ct,
            SseSource::CustomerKey {
                key: &m.key,
                key_md5: &m.key_md5,
            },
        )
        .unwrap_err();
        assert!(matches!(err, SseError::WrongCustomerKey), "got {err:?}");
    }

    #[test]
    fn s4e3_tampered_md5_with_matching_supplied_md5_fails_aead() {
        // Both stored md5 AND supplied md5 are flipped to the same bogus
        // value. The fingerprint check passes (they match) but AAD
        // authenticates the *original* md5, so AES-GCM fails.
        let m = cust_key(3);
        let mut ct = encrypt_with_source(b"x", (&m).into()).to_vec();
        ct[5] ^= 0xFF;
        let mut bogus_md5 = m.key_md5;
        bogus_md5[0] ^= 0xFF;
        let err = decrypt(
            &ct,
            SseSource::CustomerKey {
                key: &m.key,
                key_md5: &bogus_md5,
            },
        )
        .unwrap_err();
        assert!(matches!(err, SseError::DecryptFailed), "got {err:?}");
    }

    #[test]
    fn s4e3_tampered_ciphertext_fails_aead() {
        let m = cust_key(8);
        let mut ct = encrypt_with_source(b"sealed message", (&m).into()).to_vec();
        let last = ct.len() - 1;
        ct[last] ^= 0x01;
        let err = decrypt(&ct, &m).unwrap_err();
        assert!(matches!(err, SseError::DecryptFailed), "got {err:?}");
    }

    #[test]
    fn s4e3_tampered_algo_byte_rejected() {
        let m = cust_key(9);
        let mut ct = encrypt_with_source(b"x", (&m).into()).to_vec();
        ct[4] = 99;
        let err = decrypt(&ct, &m).unwrap_err();
        assert!(matches!(err, SseError::UnsupportedAlgo { tag: 99 }));
    }

    #[test]
    fn s4e3_uses_random_nonce() {
        let m = cust_key(10);
        let a = encrypt_with_source(b"deterministic input", (&m).into());
        let b = encrypt_with_source(b"deterministic input", (&m).into());
        assert_ne!(a, b, "nonce must be random per-call");
    }

    #[test]
    fn parse_customer_key_headers_happy_path() {
        let key = [11u8; KEY_LEN];
        let md5 = compute_key_md5(&key);
        let key_b64 = base64::engine::general_purpose::STANDARD.encode(key);
        let md5_b64 = base64::engine::general_purpose::STANDARD.encode(md5);
        let m = parse_customer_key_headers("AES256", &key_b64, &md5_b64).unwrap();
        assert_eq!(m.key, key);
        assert_eq!(m.key_md5, md5);
    }

    #[test]
    fn parse_customer_key_headers_rejects_wrong_algorithm() {
        let key = [1u8; KEY_LEN];
        let md5 = compute_key_md5(&key);
        let kb = base64::engine::general_purpose::STANDARD.encode(key);
        let mb = base64::engine::general_purpose::STANDARD.encode(md5);
        let err = parse_customer_key_headers("AES128", &kb, &mb).unwrap_err();
        assert!(
            matches!(err, SseError::CustomerKeyAlgorithmUnsupported { ref algo } if algo == "AES128"),
            "got {err:?}"
        );
        // Lowercase variant still rejected (AWS S3 accepts only "AES256").
        let err2 = parse_customer_key_headers("aes256", &kb, &mb).unwrap_err();
        assert!(
            matches!(err2, SseError::CustomerKeyAlgorithmUnsupported { .. }),
            "got {err2:?}"
        );
    }

    #[test]
    fn parse_customer_key_headers_rejects_wrong_key_length() {
        let short_key = vec![5u8; 16]; // half-length AES key
        let md5 = compute_key_md5(&short_key);
        let kb = base64::engine::general_purpose::STANDARD.encode(&short_key);
        let mb = base64::engine::general_purpose::STANDARD.encode(md5);
        let err = parse_customer_key_headers("AES256", &kb, &mb).unwrap_err();
        assert!(
            matches!(err, SseError::InvalidCustomerKey { reason } if reason.contains("key length")),
            "got {err:?}"
        );
    }

    #[test]
    fn parse_customer_key_headers_rejects_wrong_md5_length() {
        let key = [3u8; KEY_LEN];
        let kb = base64::engine::general_purpose::STANDARD.encode(key);
        // Truncated MD5 (15 bytes instead of 16).
        let bad_md5 = vec![0u8; 15];
        let mb = base64::engine::general_purpose::STANDARD.encode(bad_md5);
        let err = parse_customer_key_headers("AES256", &kb, &mb).unwrap_err();
        assert!(
            matches!(err, SseError::InvalidCustomerKey { reason } if reason.contains("MD5 length")),
            "got {err:?}"
        );
    }

    #[test]
    fn parse_customer_key_headers_rejects_md5_mismatch() {
        let key = [4u8; KEY_LEN];
        let other = [5u8; KEY_LEN];
        let kb = base64::engine::general_purpose::STANDARD.encode(key);
        let wrong_md5 = compute_key_md5(&other);
        let mb = base64::engine::general_purpose::STANDARD.encode(wrong_md5);
        let err = parse_customer_key_headers("AES256", &kb, &mb).unwrap_err();
        assert!(
            matches!(err, SseError::InvalidCustomerKey { reason } if reason.contains("MD5 does not match")),
            "got {err:?}"
        );
    }

    #[test]
    fn parse_customer_key_headers_rejects_bad_base64() {
        let valid_key = [0u8; KEY_LEN];
        let md5 = compute_key_md5(&valid_key);
        let mb = base64::engine::general_purpose::STANDARD.encode(md5);
        let err = parse_customer_key_headers("AES256", "!!!not-base64!!!", &mb).unwrap_err();
        assert!(
            matches!(err, SseError::InvalidCustomerKey { reason } if reason.contains("base64")),
            "got {err:?}"
        );
        // Bad MD5 base64.
        let kb = base64::engine::general_purpose::STANDARD.encode(valid_key);
        let err2 = parse_customer_key_headers("AES256", &kb, "??not-base64??").unwrap_err();
        assert!(
            matches!(err2, SseError::InvalidCustomerKey { reason } if reason.contains("base64")),
            "got {err2:?}"
        );
    }

    #[test]
    fn parse_customer_key_headers_trims_whitespace() {
        // S3 SDKs sometimes pad headers with trailing newlines.
        let key = [12u8; KEY_LEN];
        let md5 = compute_key_md5(&key);
        let kb = format!(
            "  {}\n",
            base64::engine::general_purpose::STANDARD.encode(key)
        );
        let mb = format!(
            "\t{}  ",
            base64::engine::general_purpose::STANDARD.encode(md5)
        );
        let m = parse_customer_key_headers("AES256", &kb, &mb).unwrap();
        assert_eq!(m.key, key);
    }

    // -----------------------------------------------------------------
    // Back-compat + cross-source mixing
    // -----------------------------------------------------------------

    #[test]
    fn back_compat_decrypt_s4e1_with_keyring_source() {
        let k = key32(33);
        let legacy_ct = encrypt(&k, b"v0.4 vintage object");
        let kr = SseKeyring::new(1, Arc::clone(&k));
        // Both call styles must work — `&kr` (back-compat) and
        // `SseSource::Keyring(&kr)` (explicit).
        let plain = decrypt(&legacy_ct, &kr).unwrap();
        assert_eq!(plain.as_ref(), b"v0.4 vintage object");
        let plain2 = decrypt(&legacy_ct, SseSource::Keyring(&kr)).unwrap();
        assert_eq!(plain2.as_ref(), b"v0.4 vintage object");
    }

    #[test]
    fn back_compat_decrypt_s4e2_with_keyring_source() {
        let kr = keyring_single(34);
        let ct = encrypt_v2(b"v0.5 #29 object", &kr);
        let plain = decrypt(&ct, &kr).unwrap();
        assert_eq!(plain.as_ref(), b"v0.5 #29 object");
        // encrypt_with_source(Keyring) should produce the same wire
        // format (S4E2).
        let ct2 = encrypt_with_source(b"v0.5 #29 object", SseSource::Keyring(&kr));
        assert_eq!(&ct2[..4], SSE_MAGIC_V2);
        let plain2 = decrypt(&ct2, &kr).unwrap();
        assert_eq!(plain2.as_ref(), b"v0.5 #29 object");
    }

    #[test]
    fn s4e2_blob_with_customer_key_source_is_rejected() {
        // An object stored with SSE-S4 (S4E2) but a client sending
        // SSE-C headers on the GET — this is a misuse, surface as
        // CustomerKeyUnexpected so service.rs can return 400.
        let kr = keyring_single(50);
        let ct = encrypt_v2(b"server-managed object", &kr);
        let m = cust_key(99);
        let err = decrypt(
            &ct,
            SseSource::CustomerKey {
                key: &m.key,
                key_md5: &m.key_md5,
            },
        )
        .unwrap_err();
        assert!(
            matches!(err, SseError::CustomerKeyUnexpected),
            "got {err:?}"
        );
    }

    #[test]
    fn s4e3_blob_with_keyring_source_is_rejected() {
        // Inverse: object is SSE-C (S4E3) but client forgot to send
        // SSE-C headers. Service.rs should map this to 400.
        let m = cust_key(60);
        let ct = encrypt_with_source(b"customer-key object", (&m).into());
        let kr = keyring_single(60);
        let err = decrypt(&ct, &kr).unwrap_err();
        assert!(matches!(err, SseError::CustomerKeyRequired), "got {err:?}");
    }

    #[test]
    fn looks_encrypted_detects_s4e3() {
        let m = cust_key(13);
        let ct = encrypt_with_source(b"x", (&m).into());
        assert!(looks_encrypted(&ct));
    }

    #[test]
    fn s4e3_rejects_short_body() {
        // 36 bytes passes the looks_encrypted gate but is shorter than
        // S4E3's 49-byte header.
        let mut short = Vec::new();
        short.extend_from_slice(SSE_MAGIC_V3);
        short.push(ALGO_AES_256_GCM);
        // Padding to 36 bytes (SSE_HEADER_BYTES) so the outer length
        // check passes but the S4E3 inner check fails.
        short.extend_from_slice(&[0u8; SSE_HEADER_BYTES - 5]);
        assert_eq!(short.len(), SSE_HEADER_BYTES);
        let m = cust_key(1);
        let err = decrypt(
            &short,
            SseSource::CustomerKey {
                key: &m.key,
                key_md5: &m.key_md5,
            },
        )
        .unwrap_err();
        assert!(matches!(err, SseError::TooShort { .. }), "got {err:?}");
    }

    #[test]
    fn customer_key_material_debug_redacts_key() {
        let m = cust_key(99);
        let s = format!("{m:?}");
        assert!(s.contains("redacted"));
        assert!(!s.contains(&format!("{:?}", m.key.as_slice())));
    }

    #[test]
    fn constant_time_eq_basic() {
        assert!(constant_time_eq(b"abc", b"abc"));
        assert!(!constant_time_eq(b"abc", b"abd"));
        assert!(!constant_time_eq(b"abc", b"abcd"));
        assert!(constant_time_eq(b"", b""));
    }

    #[test]
    fn compute_key_md5_known_vector() {
        // Empty input MD5 is known: d41d8cd98f00b204e9800998ecf8427e.
        let got = compute_key_md5(b"");
        let expected_hex = "d41d8cd98f00b204e9800998ecf8427e";
        assert_eq!(hex_lower(&got), expected_hex);
    }

    // -----------------------------------------------------------------
    // v0.5 #28 — SSE-KMS envelope (S4E4) tests
    // -----------------------------------------------------------------

    use crate::kms::{KmsBackend, LocalKms};
    use std::collections::HashMap;
    use std::path::PathBuf;

    fn local_kms_with(key_ids: &[(&str, [u8; 32])]) -> LocalKms {
        let mut keks: HashMap<String, [u8; 32]> = HashMap::new();
        for (id, k) in key_ids {
            keks.insert((*id).to_string(), *k);
        }
        LocalKms::from_keks(PathBuf::from("/tmp/none"), keks)
    }

    #[tokio::test]
    async fn s4e4_roundtrip_via_local_kms() {
        let kms = local_kms_with(&[("alpha", [42u8; 32])]);
        let (dek_vec, wrapped) = kms.generate_dek("alpha").await.unwrap();
        let mut dek = [0u8; 32];
        dek.copy_from_slice(&dek_vec);
        let pt = b"SSE-KMS envelope payload across the S4E4 frame";
        let ct = encrypt_with_source(
            pt,
            SseSource::Kms {
                dek: &dek,
                wrapped: &wrapped,
            },
        );
        // Frame inspection.
        assert_eq!(&ct[..4], SSE_MAGIC_V4);
        assert_eq!(ct[4], ALGO_AES_256_GCM);
        let key_id_len = ct[5] as usize;
        assert_eq!(key_id_len, "alpha".len());
        assert_eq!(&ct[6..6 + key_id_len], b"alpha");
        // peek_magic + looks_encrypted both recognise S4E4.
        assert!(looks_encrypted(&ct));
        assert_eq!(peek_magic(&ct), Some("S4E4"));
        // Async decrypt round-trip.
        let plain = decrypt_with_kms(&ct, &kms).await.unwrap();
        assert_eq!(plain.as_ref(), pt);
    }

    #[tokio::test]
    async fn s4e4_tampered_key_id_fails_aead() {
        let kms = local_kms_with(&[("alpha", [1u8; 32]), ("beta", [2u8; 32])]);
        let (dek_vec, wrapped) = kms.generate_dek("alpha").await.unwrap();
        let mut dek = [0u8; 32];
        dek.copy_from_slice(&dek_vec);
        let mut ct = encrypt_with_source(
            b"do not redirect",
            SseSource::Kms {
                dek: &dek,
                wrapped: &wrapped,
            },
        )
        .to_vec();
        // Flip the key_id from "alpha" to "betaa" by changing the
        // first byte of the key_id field. The forged id "bltha" is
        // not in the KMS, so unwrap fails with KeyNotFound surfaced
        // through KmsBackend(KmsError::KeyNotFound).
        let key_id_off = 6;
        ct[key_id_off] = b'b';
        let err = decrypt_with_kms(&ct, &kms).await.unwrap_err();
        assert!(
            matches!(
                err,
                SseError::KmsBackend(crate::kms::KmsError::UnwrapFailed { .. })
                    | SseError::KmsBackend(crate::kms::KmsError::KeyNotFound { .. })
            ),
            "got {err:?}"
        );
    }

    #[tokio::test]
    async fn s4e4_tampered_key_id_to_real_other_id_still_fails() {
        // Wrap under "alpha" but rewrite the stored key_id to "beta"
        // (which IS in the KMS). KmsBackend will try to unwrap with
        // beta's KEK and AAD = "beta", but the wrapped bytes were
        // produced with alpha's KEK + AAD = "alpha", so the local
        // KMS unwrap fails with UnwrapFailed.
        let kms = local_kms_with(&[("alpha", [1u8; 32]), ("beta", [2u8; 32])]);
        let (dek_vec, wrapped) = kms.generate_dek("alpha").await.unwrap();
        let mut dek = [0u8; 32];
        dek.copy_from_slice(&dek_vec);
        let mut ct = encrypt_with_source(
            b"redirect attempt",
            SseSource::Kms {
                dek: &dek,
                wrapped: &wrapped,
            },
        )
        .to_vec();
        // Both "alpha" and "beta" are 5 chars long so the rewrite
        // doesn't shift any other field offsets.
        let key_id_off = 6;
        ct[key_id_off..key_id_off + 5].copy_from_slice(b"beta_");
        // Trim back to 4-byte "beta" by also shrinking the length
        // prefix would change downstream offsets — instead pad the
        // forged id to keep length stable. This mirrors the realistic
        // tampering surface (attacker can flip bytes but not change
        // the on-disk layout). The KMS now sees key_id "beta_" which
        // is unknown → KeyNotFound.
        let err = decrypt_with_kms(&ct, &kms).await.unwrap_err();
        assert!(
            matches!(
                err,
                SseError::KmsBackend(crate::kms::KmsError::KeyNotFound { .. })
            ),
            "got {err:?}"
        );
    }

    #[tokio::test]
    async fn s4e4_tampered_wrapped_dek_fails_unwrap() {
        let kms = local_kms_with(&[("k", [3u8; 32])]);
        let (dek_vec, wrapped) = kms.generate_dek("k").await.unwrap();
        let mut dek = [0u8; 32];
        dek.copy_from_slice(&dek_vec);
        let mut ct = encrypt_with_source(
            b"target body",
            SseSource::Kms {
                dek: &dek,
                wrapped: &wrapped,
            },
        )
        .to_vec();
        // Locate the wrapped_dek_len + wrapped_dek field and flip a
        // byte in the middle of the wrapped DEK. AES-GCM auth on the
        // wrap fails → KmsBackend(UnwrapFailed).
        let key_id_len = ct[5] as usize;
        let wrapped_len_off = 6 + key_id_len;
        let wrapped_off = wrapped_len_off + 4;
        let mid = wrapped_off + (wrapped.ciphertext.len() / 2);
        ct[mid] ^= 0xFF;
        let err = decrypt_with_kms(&ct, &kms).await.unwrap_err();
        assert!(
            matches!(
                err,
                SseError::KmsBackend(crate::kms::KmsError::UnwrapFailed { .. })
            ),
            "got {err:?}"
        );
    }

    #[tokio::test]
    async fn s4e4_tampered_ciphertext_fails_aead() {
        let kms = local_kms_with(&[("k", [4u8; 32])]);
        let (dek_vec, wrapped) = kms.generate_dek("k").await.unwrap();
        let mut dek = [0u8; 32];
        dek.copy_from_slice(&dek_vec);
        let mut ct = encrypt_with_source(
            b"sealed body",
            SseSource::Kms {
                dek: &dek,
                wrapped: &wrapped,
            },
        )
        .to_vec();
        let last = ct.len() - 1;
        ct[last] ^= 0x01;
        let err = decrypt_with_kms(&ct, &kms).await.unwrap_err();
        assert!(matches!(err, SseError::DecryptFailed), "got {err:?}");
    }

    #[tokio::test]
    async fn s4e4_uses_random_nonce_and_dek_per_put() {
        let kms = local_kms_with(&[("k", [5u8; 32])]);
        // Two PUTs of the same plaintext under the same KEK must
        // produce different ciphertexts (fresh DEK + fresh nonce).
        let (dek1_vec, wrapped1) = kms.generate_dek("k").await.unwrap();
        let (dek2_vec, wrapped2) = kms.generate_dek("k").await.unwrap();
        let mut dek1 = [0u8; 32];
        dek1.copy_from_slice(&dek1_vec);
        let mut dek2 = [0u8; 32];
        dek2.copy_from_slice(&dek2_vec);
        let pt = b"deterministic input";
        let a = encrypt_with_source(
            pt,
            SseSource::Kms {
                dek: &dek1,
                wrapped: &wrapped1,
            },
        );
        let b = encrypt_with_source(
            pt,
            SseSource::Kms {
                dek: &dek2,
                wrapped: &wrapped2,
            },
        );
        assert_ne!(a, b);
        // Both still decrypt round-trip.
        let plain_a = decrypt_with_kms(&a, &kms).await.unwrap();
        let plain_b = decrypt_with_kms(&b, &kms).await.unwrap();
        assert_eq!(plain_a.as_ref(), pt);
        assert_eq!(plain_b.as_ref(), pt);
    }

    #[tokio::test]
    async fn s4e4_sync_decrypt_returns_kms_async_required() {
        // The whole point of KmsAsyncRequired: passing an S4E4 body
        // to the sync `decrypt` function must surface a distinct
        // error so service.rs's GET path notices the bug rather than
        // returning a generic "wrong source" 400.
        let kms = local_kms_with(&[("k", [6u8; 32])]);
        let (dek_vec, wrapped) = kms.generate_dek("k").await.unwrap();
        let mut dek = [0u8; 32];
        dek.copy_from_slice(&dek_vec);
        let ct = encrypt_with_source(
            b"async only",
            SseSource::Kms {
                dek: &dek,
                wrapped: &wrapped,
            },
        );
        // Try via Keyring source (the default sync path).
        let kr = SseKeyring::new(1, key32(0));
        let err = decrypt(&ct, &kr).unwrap_err();
        assert!(matches!(err, SseError::KmsAsyncRequired), "got {err:?}");
    }

    #[test]
    fn back_compat_s4e1_e2_e3_still_decrypt_via_sync() {
        // After adding S4E4, the sync `decrypt` path must still
        // handle every legacy frame variant unchanged.
        let k = key32(7);
        let v1 = encrypt(&k, b"v0.4 vintage");
        let kr = SseKeyring::new(1, Arc::clone(&k));
        assert_eq!(decrypt(&v1, &kr).unwrap().as_ref(), b"v0.4 vintage");

        let v2 = encrypt_v2(b"v0.5 #29 vintage", &kr);
        assert_eq!(decrypt(&v2, &kr).unwrap().as_ref(), b"v0.5 #29 vintage");

        let m = cust_key(7);
        let v3 = encrypt_with_source(b"v0.5 #27 vintage", (&m).into());
        assert_eq!(decrypt(&v3, &m).unwrap().as_ref(), b"v0.5 #27 vintage");
    }

    #[test]
    fn peek_magic_distinguishes_all_variants() {
        // S4E1 / S4E2 / S4E3 — built from real encrypts so the
        // length gate also passes.
        let k = key32(9);
        let v1 = encrypt(&k, b"x");
        assert_eq!(peek_magic(&v1), Some("S4E1"));
        let kr = SseKeyring::new(1, Arc::clone(&k));
        let v2 = encrypt_v2(b"x", &kr);
        assert_eq!(peek_magic(&v2), Some("S4E2"));
        let m = cust_key(9);
        let v3 = encrypt_with_source(b"x", (&m).into());
        assert_eq!(peek_magic(&v3), Some("S4E3"));
        // Synthetic S4E4 magic with enough trailing bytes to clear
        // the 36-byte length gate. peek_magic does NOT validate the
        // S4E4 inner header, just the magic — that's the contract
        // (cheap dispatch signal).
        let mut v4 = Vec::new();
        v4.extend_from_slice(SSE_MAGIC_V4);
        v4.extend_from_slice(&[0u8; 40]);
        assert_eq!(peek_magic(&v4), Some("S4E4"));
        // Unknown magic / too-short input → None.
        assert!(peek_magic(b"NOPE").is_none());
        assert!(peek_magic(b"short").is_none());
        assert!(peek_magic(&[0u8; 100]).is_none());
    }

    #[tokio::test]
    async fn s4e4_truncated_frame_errors_cleanly() {
        // Truncate to less than the minimum header. Must surface
        // KmsFrameTooShort, not panic, not return BadMagic.
        let truncated = b"S4E4\x01\x05hi";
        let kms = local_kms_with(&[("k", [1u8; 32])]);
        let err = decrypt_with_kms(truncated, &kms).await.unwrap_err();
        assert!(
            matches!(err, SseError::KmsFrameTooShort { .. }),
            "got {err:?}"
        );
    }

    #[tokio::test]
    async fn s4e4_oob_key_id_len_errors() {
        // Build a body that claims key_id_len = 200 but only has 4
        // bytes after the length prefix. parse_s4e4_header must
        // refuse with KmsFrameFieldOob, not slice-panic.
        let mut body = Vec::new();
        body.extend_from_slice(SSE_MAGIC_V4);
        body.push(ALGO_AES_256_GCM);
        body.push(200u8); // key_id_len
        // Remaining bytes < 200; pad to clear the looks_encrypted
        // floor (36 bytes) but stay short of the claimed key_id +
        // wrapped_dek_len + nonce + tag layout.
        body.extend_from_slice(&[0u8; 50]);
        let kms = local_kms_with(&[("k", [1u8; 32])]);
        let err = decrypt_with_kms(&body, &kms).await.unwrap_err();
        assert!(
            matches!(err, SseError::KmsFrameFieldOob { .. }),
            "got {err:?}"
        );
    }

    #[tokio::test]
    async fn s4e4_via_keyring_source_into_sync_decrypt_is_kms_async_required() {
        // S4E4 + Keyring source: sync decrypt sees the S4E4 magic
        // first and returns KmsAsyncRequired regardless of source —
        // the source mismatch never gets a chance to surface, which
        // is the right behaviour (caller's bug is "didn't peek
        // magic" not "wrong source").
        let kms = local_kms_with(&[("k", [9u8; 32])]);
        let (dek_vec, wrapped) = kms.generate_dek("k").await.unwrap();
        let mut dek = [0u8; 32];
        dek.copy_from_slice(&dek_vec);
        let ct = encrypt_with_source(
            b"x",
            SseSource::Kms {
                dek: &dek,
                wrapped: &wrapped,
            },
        );
        let m = cust_key(1);
        let err = decrypt(&ct, &m).unwrap_err();
        assert!(matches!(err, SseError::KmsAsyncRequired), "got {err:?}");
    }

    #[tokio::test]
    async fn s4e4_looks_encrypted_passthrough_returns_false_for_synthetic() {
        // S4F4 (note F not E) must NOT be confused with S4E4.
        let mut not_s4e4 = Vec::new();
        not_s4e4.extend_from_slice(b"S4F4");
        not_s4e4.extend_from_slice(&[0u8; 60]);
        assert!(!looks_encrypted(&not_s4e4));
        assert_eq!(peek_magic(&not_s4e4), None);
    }

    #[tokio::test]
    async fn s4e4_aad_length_prefix_prevents_byte_shifting() {
        // Constructing an S4E4 body where the wrapped_dek_len is
        // shrunk by N bytes and the same N bytes are prepended to
        // the key_id-equivalent area would, without length-prefixed
        // AAD, produce the same AAD bytestream. Verify our AAD
        // includes the length prefixes by tampering with
        // wrapped_dek_len and confirming AES-GCM auth fails.
        let kms = local_kms_with(&[("kk", [11u8; 32])]);
        let (dek_vec, wrapped) = kms.generate_dek("kk").await.unwrap();
        let mut dek = [0u8; 32];
        dek.copy_from_slice(&dek_vec);
        let mut ct = encrypt_with_source(
            b"length-shift defense",
            SseSource::Kms {
                dek: &dek,
                wrapped: &wrapped,
            },
        )
        .to_vec();
        let key_id_len = ct[5] as usize;
        let wrapped_len_off = 6 + key_id_len;
        // Shrink wrapped_dek_len by 1. parse_s4e4_header now reads a
        // shorter wrapped_dek and a different nonce/tag/ciphertext
        // alignment — KMS unwrap fails OR AES-GCM fails OR frame
        // bounds reject. All three surface as auditable errors;
        // none should reach a successful decrypt.
        let original_len = u32::from_be_bytes([
            ct[wrapped_len_off],
            ct[wrapped_len_off + 1],
            ct[wrapped_len_off + 2],
            ct[wrapped_len_off + 3],
        ]);
        let new_len = (original_len - 1).to_be_bytes();
        ct[wrapped_len_off..wrapped_len_off + 4].copy_from_slice(&new_len);
        let err = decrypt_with_kms(&ct, &kms).await.unwrap_err();
        // Acceptable failure modes: unwrap fail (truncated wrapped
        // DEK), AES-GCM fail (shifted nonce/tag/AAD), or frame bounds.
        assert!(
            matches!(
                err,
                SseError::KmsBackend(_)
                    | SseError::DecryptFailed
                    | SseError::KmsFrameFieldOob { .. }
                    | SseError::KmsFrameTooShort { .. }
            ),
            "got {err:?}"
        );
    }

    // -----------------------------------------------------------------------
    // v0.8 #52: S4E5 chunked SSE-S4 — encrypt_v2_chunked / decrypt_chunked_stream
    // -----------------------------------------------------------------------

    use futures::StreamExt;

    /// Drain a chunked-decrypt stream into a `Vec<Bytes>` for assertion.
    /// Surfaces the first error verbatim (so tests can match on it).
    async fn collect_chunks(
        s: impl futures::Stream<Item = Result<Bytes, SseError>>,
    ) -> Result<Vec<Bytes>, SseError> {
        let mut out = Vec::new();
        let mut s = std::pin::pin!(s);
        while let Some(item) = s.next().await {
            out.push(item?);
        }
        Ok(out)
    }

    #[test]
    fn s4e6_encrypt_layout_10mb_at_1mib() {
        // v0.8.1 #57: encrypt_v2_chunked now emits S4E6 (24-byte
        // header, 8-byte salt). 10 MB plaintext at 1 MiB chunk
        // size → magic "S4E6", chunk_count=10, header bytes line
        // up to the documented 24 + 10 * 16 + 10 MB layout.
        let kr = keyring_single(0x42);
        let chunk_size = 1024 * 1024;
        let pt_len = 10 * 1024 * 1024;
        let pt = vec![0xAB_u8; pt_len];
        let ct = encrypt_v2_chunked(&pt, &kr, chunk_size).expect("encrypt ok");
        assert_eq!(&ct[..4], SSE_MAGIC_V6, "new PUTs emit S4E6 (v0.8.1 #57)");
        assert_eq!(ct[4], ALGO_AES_256_GCM);
        assert_eq!(
            u16::from_be_bytes([ct[5], ct[6]]),
            1,
            "key_id BE = active id"
        );
        assert_eq!(ct[7], 0, "reserved must be 0");
        assert_eq!(
            u32::from_be_bytes([ct[8], ct[9], ct[10], ct[11]]),
            chunk_size as u32,
            "chunk_size BE",
        );
        assert_eq!(
            u32::from_be_bytes([ct[12], ct[13], ct[14], ct[15]]),
            10,
            "chunk_count BE — 10 MiB / 1 MiB = 10 (no remainder)",
        );
        // Salt now 8 bytes — verify the slice exists and isn't all
        // zeros (defensive: catches a stuck PRNG that would leave
        // the salt array uninitialized).
        assert_eq!(&ct[16..24].len(), &8, "S4E6 salt slot is 8 bytes");
        assert_ne!(
            &ct[16..24],
            &[0u8; 8],
            "S4E6 salt must be random, not zeros"
        );
        assert_eq!(
            ct.len(),
            S4E6_HEADER_BYTES + 10 * S4E6_PER_CHUNK_OVERHEAD + pt_len,
            "total = header (24) + 10 tags + plaintext",
        );
        assert!(looks_encrypted(&ct), "looks_encrypted must accept S4E6");
        assert_eq!(peek_magic(&ct), Some("S4E6"));
    }

    #[tokio::test]
    async fn s4e6_decrypt_chunked_stream_byte_equal() {
        // v0.8.1 #57: round-trip via S4E6 path. encrypt_v2_chunked
        // emits S4E6, decrypt_chunked_stream consumes S4E5/S4E6.
        let kr = keyring_single(0x55);
        let pt: Vec<u8> = (0..(10 * 1024 * 1024_u32))
            .map(|i| (i & 0xFF) as u8)
            .collect();
        let ct = encrypt_v2_chunked(&pt, &kr, 1024 * 1024).unwrap();
        // Sanity: the new PUT is S4E6.
        assert_eq!(&ct[..4], SSE_MAGIC_V6, "new emit is S4E6");
        let stream = decrypt_chunked_stream(ct, &kr);
        let chunks = collect_chunks(stream).await.expect("stream ok");
        assert_eq!(chunks.len(), 10, "10 chunks expected for 10 MiB / 1 MiB");
        let mut joined = Vec::with_capacity(pt.len());
        for c in chunks {
            joined.extend_from_slice(&c);
        }
        assert_eq!(joined.len(), pt.len(), "byte length matches");
        assert_eq!(joined, pt, "byte-equal round-trip");
    }

    #[tokio::test]
    async fn s4e6_single_chunk_for_small_object() {
        // Plaintext smaller than chunk_size → chunk_count=1. The
        // chunk_count field offset is unchanged between S4E5 and
        // S4E6 (both at body[12..16]); only the salt width differs.
        let kr = keyring_single(0x77);
        let pt = b"tiny payload, smaller than chunk_size";
        let ct = encrypt_v2_chunked(pt, &kr, 1024 * 1024).unwrap();
        assert_eq!(
            u32::from_be_bytes([ct[12], ct[13], ct[14], ct[15]]),
            1,
            "small plaintext = single chunk",
        );
        let stream = decrypt_chunked_stream(ct, &kr);
        let chunks = collect_chunks(stream).await.expect("stream ok");
        assert_eq!(chunks.len(), 1);
        assert_eq!(chunks[0].as_ref(), pt);
    }

    #[tokio::test]
    async fn s4e6_tampered_chunk_n_reports_chunk_index() {
        // v0.8.1 #57: same tamper-detect contract as the S4E5
        // version, just with the wider 24-byte header. Tamper
        // byte inside chunk index 3 (= 4th chunk) — the stream
        // must yield 3 successful chunks, then ChunkAuthFailed { 3 }.
        let kr = keyring_single(0x91);
        let chunk_size = 1024;
        let pt = vec![0xCD_u8; chunk_size * 8]; // 8 chunks
        let mut ct = encrypt_v2_chunked(&pt, &kr, chunk_size).unwrap().to_vec();
        // Locate chunk 3's first ciphertext byte: header (24) + 3 *
        // (tag 16 + ct 1024) + tag 16 = 24 + 3*1040 + 16 = 3160.
        let target = S4E6_HEADER_BYTES + 3 * (TAG_LEN + chunk_size) + TAG_LEN;
        ct[target] ^= 0x42;
        let stream = decrypt_chunked_stream(bytes::Bytes::from(ct), &kr);
        let mut s = std::pin::pin!(stream);
        // Chunks 0, 1, 2 must succeed.
        for expected_i in 0..3_u32 {
            let item = s.next().await.expect("yield");
            item.unwrap_or_else(|e| panic!("chunk {expected_i}: {e:?}"));
        }
        // Chunk 3 fails with the right index.
        let err = s.next().await.expect("yield error").unwrap_err();
        assert!(
            matches!(err, SseError::ChunkAuthFailed { chunk_index: 3 }),
            "got {err:?}",
        );
    }

    #[tokio::test]
    async fn s4e5_back_compat_s4e2_blob_rejected_with_clear_error() {
        // Feeding an S4E2 frame to decrypt_chunked_stream should
        // surface BadMagic on the first poll (NOT silently fall
        // back — the caller is expected to peek_magic and dispatch).
        let kr = keyring_single(0x12);
        let s4e2 = encrypt_v2(b"a v2 blob, not chunked", &kr);
        let stream = decrypt_chunked_stream(s4e2, &kr);
        let result = collect_chunks(stream).await;
        let err = result.unwrap_err();
        assert!(matches!(err, SseError::BadMagic { .. }), "got {err:?}");
    }

    #[test]
    fn s4e6_salt_randomness_smoke() {
        // 8-byte salt → birthday paradox 50% collision at ~2^32
        // (~4.3 billion) PUTs. 1024 PUTs → effectively zero
        // expected collisions; we don't enforce zero, just
        // sanity-check the salt actually differs more than half
        // the time (catches a stuck PRNG without a 4-billion-PUT
        // test).
        let kr = keyring_single(0x33);
        let mut salts = std::collections::HashSet::new();
        let n = 1024;
        for _ in 0..n {
            let ct = encrypt_v2_chunked(b"x", &kr, 64).unwrap();
            let mut salt = [0u8; 8];
            salt.copy_from_slice(&ct[16..24]);
            salts.insert(salt);
        }
        assert!(
            salts.len() > n / 2,
            "expected most of the {n} salts to be unique (got {} unique)",
            salts.len(),
        );
    }

    #[test]
    fn s4e6_chunk_size_zero_invalid() {
        let kr = keyring_single(0x66);
        let err = encrypt_v2_chunked(b"hi", &kr, 0).unwrap_err();
        assert!(matches!(err, SseError::ChunkSizeInvalid));
    }

    #[tokio::test]
    async fn s4e6_truncated_body_reports_frame_truncated() {
        // Truncate inside chunk 2's tag → ChunkFrameTruncated, not
        // panic, not silent success. Header is now 24 bytes (S4E6).
        let kr = keyring_single(0xA1);
        let chunk_size = 256;
        let pt = vec![0u8; chunk_size * 4];
        let ct = encrypt_v2_chunked(&pt, &kr, chunk_size).unwrap();
        // Truncate to inside chunk 2's tag: header + chunk0 + chunk1
        // + 8B partial of chunk2's tag.
        let trunc = S4E6_HEADER_BYTES + 2 * (TAG_LEN + chunk_size) + 8;
        let truncated = bytes::Bytes::copy_from_slice(&ct[..trunc]);
        let stream = decrypt_chunked_stream(truncated, &kr);
        let result = collect_chunks(stream).await;
        let err = result.unwrap_err();
        assert!(
            matches!(err, SseError::ChunkFrameTruncated { .. }),
            "got {err:?}",
        );
    }

    #[test]
    fn s4e6_decrypt_buffered_round_trip_via_top_level_decrypt() {
        // Sync `decrypt(blob, &keyring)` must also accept the
        // chunked frames (back-compat path for callers that need
        // the whole plaintext).
        let kr = keyring_single(0xDE);
        let pt = b"buffered sync decrypt path".repeat(32);
        let ct = encrypt_v2_chunked(&pt, &kr, 13).unwrap();
        let plain = decrypt(&ct, &kr).expect("buffered S4E6 decrypt ok");
        assert_eq!(plain.as_ref(), pt.as_slice());
    }

    #[tokio::test]
    async fn s4e6_unknown_key_id_in_frame_errors() {
        // Encrypt under id=7, decrypt under a keyring that lacks id=7.
        let kr_put = SseKeyring::new(7, key32(0xCC));
        let kr_get = keyring_single(0xCC); // only id=1
        let ct = encrypt_v2_chunked(b"orphan key", &kr_put, 64).unwrap();
        // Sync path
        let err = decrypt(&ct, &kr_get).unwrap_err();
        assert!(
            matches!(err, SseError::KeyNotInKeyring { id: 7 }),
            "got {err:?}"
        );
        // Stream path
        let stream = decrypt_chunked_stream(ct, &kr_get);
        let result = collect_chunks(stream).await;
        assert!(
            matches!(result, Err(SseError::KeyNotInKeyring { id: 7 })),
            "got {result:?}",
        );
    }

    #[tokio::test]
    async fn s4e6_final_chunk_smaller_than_chunk_size() {
        // Plaintext = 2.5 chunks → final chunk holds half the bytes.
        // S4E6 header = 24 bytes → total on-disk = 24 + 48 + 250.
        let kr = keyring_single(0xEF);
        let chunk_size = 100;
        let pt: Vec<u8> = (0..250_u32).map(|i| i as u8).collect();
        let ct = encrypt_v2_chunked(&pt, &kr, chunk_size).unwrap();
        assert_eq!(
            u32::from_be_bytes([ct[12], ct[13], ct[14], ct[15]]),
            3,
            "ceil(250/100) = 3 chunks",
        );
        // Total on-disk: 24 header + 3 tags (48) + 250 plaintext = 322.
        assert_eq!(ct.len(), S4E6_HEADER_BYTES + 48 + 250);
        let stream = decrypt_chunked_stream(ct, &kr);
        let chunks = collect_chunks(stream).await.expect("stream ok");
        assert_eq!(chunks.len(), 3);
        assert_eq!(chunks[0].len(), 100);
        assert_eq!(chunks[1].len(), 100);
        assert_eq!(chunks[2].len(), 50, "final chunk is the remainder");
        let joined: Vec<u8> = chunks.iter().flat_map(|c| c.iter().copied()).collect();
        assert_eq!(joined, pt);
    }

    // -----------------------------------------------------------------------
    // v0.8.1 #57: S4E6-specific tests added on top of the renamed
    // s4e6_* battery above. Keep these focused on what's *new*:
    //   - back-compat read of legacy S4E5 blobs
    //   - 24-bit chunk_count cap
    //   - the parse_s4e6_header public API
    // -----------------------------------------------------------------------

    #[test]
    fn s4e6_back_compat_read_s4e5_blob() {
        // Synthesize a "v0.8.0 vintage" S4E5 blob via the test-only
        // helper, then prove the v0.8.1 gateway decrypts it under
        // the same keyring — both via sync `decrypt` (buffered)
        // and the streaming path. Without this, every S4E5 object
        // in production becomes unreadable after the upgrade.
        let kr = keyring_single(0x57);
        let pt = b"v0.8.0 vintage chunked SSE-S4 object".repeat(64);
        let s4e5 = encrypt_v2_chunked_s4e5_for_test(&pt, &kr, 91).unwrap();
        // Confirm the test fixture really is S4E5 magic + 20-byte header.
        assert_eq!(&s4e5[..4], SSE_MAGIC_V5, "fixture must be S4E5");
        assert_eq!(peek_magic(&s4e5), Some("S4E5"));
        // Sync decrypt path (top-level `decrypt`, dispatches V5 + V6).
        let plain_sync = decrypt(&s4e5, &kr).expect("sync S4E5 decrypt ok");
        assert_eq!(plain_sync.as_ref(), pt.as_slice());
        // Streaming decrypt path — must also accept S4E5.
        let collected = futures::executor::block_on(async {
            let stream = decrypt_chunked_stream(s4e5.clone(), &kr);
            collect_chunks(stream).await
        })
        .expect("stream S4E5 decrypt ok");
        let mut joined = Vec::with_capacity(pt.len());
        for c in collected {
            joined.extend_from_slice(&c);
        }
        assert_eq!(joined, pt, "S4E5 streaming round-trip byte-equal");
    }

    #[test]
    fn s4e6_layout_24_bytes_header() {
        // Sanity check: the S4E6 fixed header is exactly 24 bytes
        // (vs 20 for S4E5). Catches an accidental const drift in a
        // future PR.
        assert_eq!(S4E6_HEADER_BYTES, 24);
        assert_eq!(S4E6_PER_CHUNK_OVERHEAD, TAG_LEN);
        assert_eq!(S4E6_HEADER_BYTES, S4E5_HEADER_BYTES + 4);
    }

    #[test]
    fn s4e6_parse_header_round_trip() {
        // parse_s4e6_header is the public mirror of the internal
        // parse_chunked_header, useful for admin tools. Verify it
        // returns the same field values that encrypt_v2_chunked wrote.
        let kr = keyring_single(0xAB);
        let chunk_size = 256;
        let pt = vec![1u8; 7 * chunk_size];
        let ct = encrypt_v2_chunked(&pt, &kr, chunk_size).unwrap();
        let hdr = parse_s4e6_header(&ct).expect("parse ok");
        assert_eq!(hdr.key_id, 1);
        assert_eq!(hdr.chunk_size, chunk_size as u32);
        assert_eq!(hdr.chunk_count, 7);
        assert_eq!(hdr.salt.len(), 8);
        // Bad magic on a non-S4E6 blob → BadMagic.
        let bogus = b"S4E2\x01\x00\x00\x00........................";
        let err = parse_s4e6_header(bogus).unwrap_err();
        assert!(matches!(err, SseError::BadMagic { .. }), "got {err:?}");
        // Truncation → ChunkFrameTruncated.
        let err2 = parse_s4e6_header(&ct[..10]).unwrap_err();
        assert!(
            matches!(err2, SseError::ChunkFrameTruncated { .. }),
            "got {err2:?}"
        );
    }

    #[test]
    fn s4e6_salt_uniqueness_smoke_16m() {
        // v0.8.1 #57 security regression detection: with the
        // 4-byte S4E5 salt, ~65,536 PUTs already had ~50%
        // birthday collision (the bug that motivated this patch).
        // With the 8-byte S4E6 salt the expected collisions over
        // 65,536 PUTs is ~2^16 * 2^16 / 2^65 ≈ 2^-33 — i.e.
        // effectively zero.
        //
        // We can't actually run 16M PUTs in unit-test wall-clock
        // (each PUT does an AES-GCM encrypt), so we run a fast
        // smoke (16k) and additionally validate the math: at the
        // S4E5 4-byte salt width, 16k PUTs would already give a
        // ~3.1% collision probability by birthday bound; at the
        // S4E6 8-byte salt that drops to ~3.6e-11. The smoke test
        // therefore *would* show collisions if we'd accidentally
        // shipped the 4-byte salt — confirming the regression
        // detector.
        let kr = keyring_single(0xA6);
        let mut salts = std::collections::HashSet::with_capacity(16384);
        let n = 16384_usize;
        let mut collisions_top4 = 0usize;
        let mut top4_seen = std::collections::HashSet::with_capacity(16384);
        for _ in 0..n {
            let ct = encrypt_v2_chunked(b"x", &kr, 64).unwrap();
            let mut salt = [0u8; 8];
            salt.copy_from_slice(&ct[16..24]);
            salts.insert(salt);
            // Side-channel: count collisions on just the *first 4
            // bytes* of the 8-byte salt. If we'd kept the old
            // 4-byte salt, this collision count would be the only
            // collision count — and at n=16k it should be ~62
            // (birthday: n^2/(2 * 2^32) = 16384^2/2^33 ≈ 31, with
            // some noise). The full 8-byte salt test passes if the
            // FULL salts are all unique while the truncated-to-4
            // count is non-zero, proving the extra 4 bytes really
            // are doing the security work.
            let mut top4 = [0u8; 4];
            top4.copy_from_slice(&salt[..4]);
            if !top4_seen.insert(top4) {
                collisions_top4 += 1;
            }
        }
        assert_eq!(
            salts.len(),
            n,
            "all 8-byte salts must be unique across {n} PUTs (got {} unique)",
            salts.len(),
        );
        // Sanity check the regression detector: at 16k PUTs with a
        // 4-byte salt, birthday math predicts ~31 collisions on
        // average. Anything in the 0..200 range is statistically
        // believable for 16k uniform 32-bit draws; we only assert
        // ">= 1" (i.e. at least one collision happened — which
        // would have been a real bug under S4E5).
        eprintln!(
            "s4e6_salt_uniqueness_smoke_16m: 16k PUTs, full 8B salts \
             all unique ({}/{}), simulated 4B-truncated salt yielded \
             {} collisions (this is what S4E5 would have shipped)",
            salts.len(),
            n,
            collisions_top4,
        );
        // Don't make the test flaky on the simulated number (it's a
        // statistical signal); just leave the eprintln for the
        // operator audit log when the test runs verbose.
    }

    #[test]
    fn s4e6_max_chunks_24bit() {
        // The S4E6 nonce embeds the chunk index as 24-bit BE, so
        // chunk_count > 2^24 - 1 must surface ChunkCountTooLarge
        // at PUT time. We can't actually run a 16M-chunk encrypt
        // in unit-test wall-clock (16M AES-GCM tag verifies even
        // on AES-NI is several minutes), but we can verify the
        // CAP constant matches expectations + exercise the cap by
        // picking a chunk_size that forces overflow on a tiny
        // plaintext.
        assert_eq!(S4E6_MAX_CHUNK_COUNT, (1u32 << 24) - 1);
        assert_eq!(S4E6_MAX_CHUNK_COUNT, 16_777_215);

        // chunk_size=1 + plaintext.len()=16_777_216 → 16M+1 chunks
        // → over the cap → ChunkCountTooLarge. Allocating a 16
        // MiB plaintext is fine.
        let kr = keyring_single(0xC4);
        let pt = vec![0u8; (S4E6_MAX_CHUNK_COUNT as usize) + 1]; // 16,777,216 B
        let err = encrypt_v2_chunked(&pt, &kr, 1).unwrap_err();
        assert!(
            matches!(
                err,
                SseError::ChunkCountTooLarge {
                    got: 16_777_216,
                    max: 16_777_215
                }
            ),
            "got {err:?}",
        );

        // And just under the cap (chunk_count = 16_777_215) should
        // succeed. We pick chunk_size that produces exactly the cap
        // so the inner loop only runs N times. 16M chunk-encrypts
        // would be slow, so test with a smaller cap-near config
        // that exercises the same boundary check: 1023 chunks of 1
        // byte each = 1023 chunks well under the cap → success.
        // The actual on-cap encrypt is exercised by the buffered
        // decrypt path through `parse_chunked_header`.
        let pt_ok = vec![0u8; 1023];
        let ct = encrypt_v2_chunked(&pt_ok, &kr, 1).expect("under-cap PUT must succeed");
        let hdr = parse_s4e6_header(&ct).unwrap();
        assert_eq!(hdr.chunk_count, 1023);

        // Synthesize a frame that *claims* chunk_count > cap and
        // verify the parser rejects it (defensive: a tampered
        // header should not loop the walker 16M+ times).
        let mut tampered = ct.to_vec();
        // Rewrite chunk_count BE to S4E6_MAX_CHUNK_COUNT + 1 = 2^24.
        let bad = (S4E6_MAX_CHUNK_COUNT + 1).to_be_bytes();
        tampered[12..16].copy_from_slice(&bad);
        let err2 = parse_s4e6_header(&tampered).unwrap_err();
        assert!(
            matches!(
                err2,
                SseError::ChunkCountTooLarge {
                    got: 16_777_216,
                    max: 16_777_215
                }
            ),
            "got {err2:?}",
        );
    }

    #[test]
    fn s4e6_nonce_v6_layout() {
        // Direct unit test on nonce_v6: prefix b'E', then 8B salt,
        // then 24-bit BE chunk_index. The high byte of u32
        // chunk_index must be dropped (caller-validated cap).
        let salt = [0xAA_u8; 8];
        let n0 = nonce_v6(&salt, 0);
        assert_eq!(n0[0], b'E');
        assert_eq!(&n0[1..9], &salt);
        assert_eq!(&n0[9..12], &[0, 0, 0]);
        let n1 = nonce_v6(&salt, 1);
        assert_eq!(&n1[9..12], &[0, 0, 1]);
        let n_mid = nonce_v6(&salt, 0x123456);
        assert_eq!(&n_mid[9..12], &[0x12, 0x34, 0x56]);
        let n_max = nonce_v6(&salt, S4E6_MAX_CHUNK_COUNT);
        assert_eq!(&n_max[9..12], &[0xFF, 0xFF, 0xFF]);
    }

    #[tokio::test]
    async fn s4e6_tampered_salt_byte_fails_aead() {
        // Flipping a single byte of the 8-byte salt in the header
        // must invalidate every chunk's AES-GCM tag (salt is in
        // the AAD). Confirms the salt expansion didn't drop
        // header authentication.
        let kr = keyring_single(0xB6);
        let pt = b"salt-in-aad coverage".repeat(64);
        let mut ct = encrypt_v2_chunked(&pt, &kr, 128).unwrap().to_vec();
        // Salt bytes 16..24 — flip the middle byte.
        ct[20] ^= 0x01;
        let err = decrypt(&ct, &kr).unwrap_err();
        assert!(
            matches!(err, SseError::ChunkAuthFailed { chunk_index: 0 }),
            "got {err:?}",
        );
    }

    // -----------------------------------------------------------------------
    // v0.8.2 #64: pre-allocation guard for chunked SSE frames
    //
    // The buffered decrypt path used to call `Vec::with_capacity(chunk_size
    // * chunk_count)` *before* validating either factor. A 24-byte malicious
    // S4E6 header could declare a multi-PB plaintext and trigger an instant
    // OOM / panic on a tiny ciphertext. Verify the guard rejects every
    // adversarial header pattern before any allocation happens.
    // -----------------------------------------------------------------------

    /// Build a minimal S4E6 header with attacker-controlled chunk_size /
    /// chunk_count (no chunks attached — the body is exactly 24 bytes).
    /// Used by the DoS tests below to synthesize malicious frames without
    /// running an encrypt.
    fn synth_s4e6_header(chunk_size: u32, chunk_count: u32) -> Vec<u8> {
        let mut blob = Vec::with_capacity(S4E6_HEADER_BYTES);
        blob.extend_from_slice(SSE_MAGIC_V6);
        blob.push(ALGO_AES_256_GCM);
        blob.extend_from_slice(&1_u16.to_be_bytes()); // key_id = 1
        blob.push(0); // reserved
        blob.extend_from_slice(&chunk_size.to_be_bytes());
        blob.extend_from_slice(&chunk_count.to_be_bytes());
        blob.extend_from_slice(&[0u8; 8]); // 8-byte salt
        debug_assert_eq!(blob.len(), S4E6_HEADER_BYTES);
        blob
    }

    #[test]
    fn s4e6_header_claims_huge_size_rejected_pre_alloc() {
        // 1 GiB chunk_size × 100 chunks = 100 GiB declared plaintext, but
        // the body the attacker actually shipped is only 100 bytes. The
        // pre-alloc guard's cap arm fires (100 GiB > 5 GiB default cap)
        // *before* the walker / `Vec::with_capacity(100 GiB)` — surfaces
        // as ChunkFrameTooLarge.
        let kr = keyring_single(0x01);
        let chunk_size: u32 = 1 << 30; // 1 GiB
        let chunk_count: u32 = 100;
        let mut blob = synth_s4e6_header(chunk_size, chunk_count);
        // Pad with 100 random bytes — the attack model: tiny payload,
        // monster header.
        blob.extend_from_slice(&[0u8; 100]);
        let err = decrypt_chunked_buffered_default(&blob, &kr).unwrap_err();
        assert!(
            matches!(err, SseError::ChunkFrameTooLarge { .. }),
            "expected ChunkFrameTooLarge (declared 100 GiB > 5 GiB cap), got {err:?}",
        );
        // Same input, tighter caller-supplied cap (1 MiB): still
        // ChunkFrameTooLarge. No panic, no OOM either way.
        let err2 = decrypt_chunked_buffered(&blob, &kr, 1024 * 1024).unwrap_err();
        assert!(
            matches!(err2, SseError::ChunkFrameTooLarge { .. }),
            "expected ChunkFrameTooLarge under tighter cap, got {err2:?}",
        );
    }

    #[test]
    fn s4e6_header_chunk_size_x_chunk_count_overflows_u64() {
        // chunk_size = u32::MAX, chunk_count > 1 → product > u64 cap?
        // No — u32::MAX * u32::MAX fits in u64 (~2^64). To force u64
        // overflow we'd need both > 2^32. But chunk_count is capped to
        // 16M (S4E6_MAX_CHUNK_COUNT) by the earlier check, so the
        // realistic adversarial maximum is u32::MAX * 16M ≈ 2^56 — well
        // under u64::MAX. The overflow path is therefore *theoretically*
        // unreachable on S4E6 (24-bit chunk_count cap protects it), but
        // we still test it via S4E5 which has no 24-bit cap on
        // chunk_count: a 4-byte salt frame with chunk_size = u32::MAX
        // and chunk_count = u32::MAX gives 2^64 - 2^33 + 1 — fits u64
        // but adding the 16-byte tag overhead × u32::MAX chunks
        // overflows. Verify ChunkFrameTooLarge surfaces.
        let kr = keyring_single(0x02);
        // Synthesize an S4E5 header (20 bytes) with maximally
        // adversarial chunk_size + chunk_count.
        let mut blob = Vec::with_capacity(S4E5_HEADER_BYTES);
        blob.extend_from_slice(SSE_MAGIC_V5);
        blob.push(ALGO_AES_256_GCM);
        blob.extend_from_slice(&1_u16.to_be_bytes());
        blob.push(0);
        blob.extend_from_slice(&u32::MAX.to_be_bytes()); // chunk_size = 2^32 - 1
        blob.extend_from_slice(&u32::MAX.to_be_bytes()); // chunk_count = 2^32 - 1
        blob.extend_from_slice(&[0u8; 4]); // 4-byte S4E5 salt
        debug_assert_eq!(blob.len(), S4E5_HEADER_BYTES);
        let err = decrypt_chunked_buffered_default(&blob, &kr).unwrap_err();
        assert!(
            matches!(err, SseError::ChunkFrameTooLarge { .. }),
            "expected ChunkFrameTooLarge for u64 overflow, got {err:?}",
        );

        // Also smoke the same input through `parse_chunked_header`
        // directly (the streaming path) — must error, never panic.
        let direct = parse_chunked_header(&blob, usize::MAX).unwrap_err();
        assert!(
            matches!(direct, SseError::ChunkFrameTooLarge { .. }),
            "streaming path: expected ChunkFrameTooLarge, got {direct:?}",
        );
    }

    #[test]
    fn s4e6_header_within_max_body_bytes_passes() {
        // 1 MiB chunk_size × 100 chunks = 100 MiB declared plaintext —
        // well under the 5 GiB default cap. The header validation must
        // NOT reject this; instead it falls through to chunk-walk +
        // AES-GCM verify (which then fails on this synthetic frame
        // because we didn't write any ciphertext, so we expect
        // ChunkFrameTruncated *not* ChunkFrameTooLarge).
        let kr = keyring_single(0x03);
        let chunk_size: u32 = 1024 * 1024; // 1 MiB
        let chunk_count: u32 = 100;
        let mut blob = synth_s4e6_header(chunk_size, chunk_count);
        // Append the full declared chunk array so the truncation arm
        // doesn't fire — 100 chunks × (16 B tag + 1 MiB ct) = 100 MiB
        // + 1.6 KiB tags. We write zeros; AES-GCM verify will fail on
        // chunk 0, but that proves the pre-alloc guard *let it through*.
        let chunk_array_size =
            (chunk_count as usize) * (S4E6_PER_CHUNK_OVERHEAD + chunk_size as usize);
        blob.resize(blob.len() + chunk_array_size, 0);
        let err = decrypt_chunked_buffered(&blob, &kr, DEFAULT_MAX_BODY_BYTES).unwrap_err();
        // The guard let it through (we got past parse_chunked_header)
        // and AES-GCM tag verify failed on chunk 0 — that's the right
        // failure mode. ChunkFrameTooLarge would mean the cap fired
        // (a regression of this test). KeyNotInKeyring would mean the
        // header parsed but the key id (1) wasn't present (also a
        // regression — the keyring has id=1 active). Anything else is
        // a guard-too-strict bug.
        assert!(
            matches!(err, SseError::ChunkAuthFailed { chunk_index: 0 }),
            "expected ChunkAuthFailed (guard let it through), got {err:?}",
        );
    }

    #[test]
    fn s4e6_header_exceeds_max_body_bytes_rejected() {
        // 1 MiB × 6000 chunks = 6 GiB declared plaintext > 5 GiB cap.
        // Must reject as ChunkFrameTooLarge before any alloc. We don't
        // attach the 6 GiB chunk array to the body — the cap arm only
        // looks at the declared `chunk_size × chunk_count`, not the
        // actual body length, so a tiny body suffices to exercise the
        // cap.
        let kr = keyring_single(0x04);
        let chunk_size: u32 = 1024 * 1024; // 1 MiB
        let chunk_count: u32 = 6000;
        let blob = synth_s4e6_header(chunk_size, chunk_count);
        // Body is exactly the 24-byte header, no chunks attached.
        // body.len() = 24 ≤ max_total (~6 GiB) so the trailing-bytes
        // check does NOT fire; the 6 GiB > 5 GiB cap check does.
        let err = decrypt_chunked_buffered(&blob, &kr, DEFAULT_MAX_BODY_BYTES).unwrap_err();
        assert!(
            matches!(err, SseError::ChunkFrameTooLarge { .. }),
            "expected ChunkFrameTooLarge (6 GiB declared > 5 GiB cap), got {err:?}",
        );

        // Directly exercise the cap arm with a tighter cap (1 MiB)
        // and a 100 MiB declared plaintext that fits fully in the
        // body, so the cap is unambiguously the deciding factor.
        let chunk_size_b: u32 = 1024 * 1024; // 1 MiB
        let chunk_count_b: u32 = 100;
        let mut blob_b = synth_s4e6_header(chunk_size_b, chunk_count_b);
        let pad_b = (chunk_count_b as usize) * (S4E6_PER_CHUNK_OVERHEAD + chunk_size_b as usize);
        blob_b.resize(blob_b.len() + pad_b, 0);
        // Cap = 1 MiB, declared plaintext = 100 MiB → ChunkFrameTooLarge.
        let err_b = decrypt_chunked_buffered(&blob_b, &kr, 1024 * 1024).unwrap_err();
        assert!(
            matches!(err_b, SseError::ChunkFrameTooLarge { .. }),
            "expected ChunkFrameTooLarge (cap < declared), got {err_b:?}",
        );
    }

    #[test]
    fn s4e6_random_header_never_panics() {
        // DoS regression — feed 100k random byte sequences shaped like
        // chunked SSE headers to `parse_chunked_header`. Every result
        // must be Ok or Err; no panic, no OOM. This exercises the
        // arithmetic-overflow + truncation arms of the v0.8.2 #64
        // guard against adversarial inputs the previous unit tests
        // didn't enumerate.
        //
        // We use a fixed seed (deterministic) rather than proptest
        // here so the test is repeatable across CI runs without
        // pulling in shrink machinery for a pure no-panic property.
        use rand::{Rng, SeedableRng, rngs::StdRng};
        let mut rng = StdRng::seed_from_u64(0xC0FF_EE64_6464_64DE);
        let mut max_body_bytes_choices = [
            0_usize,
            1024,
            1024 * 1024,
            DEFAULT_MAX_BODY_BYTES,
            usize::MAX,
        ]
        .iter()
        .copied()
        .cycle();
        for _ in 0..100_000 {
            // Body length in [0, 256] — tiny on purpose: most random
            // bytes won't be a valid header, but we want the parser
            // to robustly reject every shape (truncated, wrong magic,
            // wrong algo, declared-vs-actual mismatch, overflow).
            let body_len = rng.gen_range(0..=256_usize);
            let mut body = vec![0u8; body_len];
            rng.fill(body.as_mut_slice());
            // Bias 1/4 of trials toward valid magic so the deeper
            // arithmetic paths get exercised, not just the BadMagic
            // early-out.
            if body_len >= 4 && rng.gen_bool(0.25) {
                if rng.gen_bool(0.5) {
                    body[..4].copy_from_slice(SSE_MAGIC_V5);
                } else {
                    body[..4].copy_from_slice(SSE_MAGIC_V6);
                }
            }
            let max_cap = max_body_bytes_choices.next().unwrap();
            // The contract: never panic, never alloc more than
            // `max_cap` worth of memory. We assert only no-panic
            // here; OOM behavior is implicit (the function returns
            // Err before any large alloc).
            let _ = parse_chunked_header(&body, max_cap);
        }
    }

    // Bonus: an extreme-overflow case the guard is specifically
    // hardened against — chunk_count = u32::MAX with the per-chunk
    // overhead multiplication forced to overflow u64. Catches a
    // future refactor that would replace `checked_mul` with `*`.
    #[test]
    fn s4e5_extreme_overflow_chunk_count_u32_max() {
        let kr = keyring_single(0x05);
        // u32::MAX chunk_size × u32::MAX chunk_count overflows u64
        // when adding the 16 * u32::MAX tag overhead. (Strictly:
        // chunk_size × chunk_count = (2^32-1)^2 ≈ 2^64 - 2^33 + 1
        // — already u64-saturating; the +tag step then overflows.)
        // The guard's `checked_add` chain must catch this.
        let mut blob = Vec::with_capacity(S4E5_HEADER_BYTES);
        blob.extend_from_slice(SSE_MAGIC_V5);
        blob.push(ALGO_AES_256_GCM);
        blob.extend_from_slice(&1_u16.to_be_bytes());
        blob.push(0);
        blob.extend_from_slice(&u32::MAX.to_be_bytes());
        blob.extend_from_slice(&u32::MAX.to_be_bytes());
        blob.extend_from_slice(&[0u8; 4]);
        let err = decrypt_chunked_buffered_default(&blob, &kr).unwrap_err();
        assert!(
            matches!(err, SseError::ChunkFrameTooLarge { .. }),
            "expected ChunkFrameTooLarge for extreme overflow, got {err:?}",
        );
    }
}