Round 4: fix 26 defects across crypto, resource use and codec paths

Twenty-six confirmed findings from the fourth audit round, landed as one
cluster because they were found by agents working over disjoint file sets.

The one worth calling out is a pair of AACS tests that could not fail.
Both asserted CBC behaviour against a hand-rolled expectation that
happened to be IV-independent, so replacing AACS_IV with sixteen zero
bytes left them passing — they were pinning the code's own arithmetic,
not the published constant. Replaced with a literal witness of the
published IV plus the NIST SP 800-38A F.2.2 CBC-AES128 vector, and
verified the other way round: zeroing AACS_IV now fails three tests.

The rest are allocation and correctness work on hot paths: the Annex-B
writer in demux_sink allocated and freed a whole-frame Vec per frame,
which for a UHD title is ~200,000 allocations over the mmap threshold
plus the page faults to first-touch each one; it now reuses a buffer on
the writer, and still takes the NAL prefix width from the configuration
record rather than assuming four.

Six findings whose real fix lives in a consumer crate are recorded for
re-filing rather than patched here.

src/aacs/content.rs

@@ -147,8 +147,14 @@ pub fn aacs_unit_needs_decrypt(unit: &[u8], format: crate::disc::ContentFormat)
 }
 
 /// Minimum synced content packets that PROVE a key opened a unit. Four `0x47`
-/// syncs ≈ 32 bits of MPEG-TS structure ≈ 1-in-4-billion that a wrong key (uniform
+/// syncs are 32 bits of MPEG-TS structure, but the per-UNIT false-pass risk is
-/// AES noise, `0x47` at 1/256 per packet) fakes it. It is an ABSOLUTE proof floor,
+/// NOT 2^-32: `is_clean_ts` accepts ANY four of the ~31 encrypted packets in a
+/// 6144-byte aligned unit, so for a wrong key (uniform AES noise, `0x47` at 1/256
+/// per packet) it is ≈ C(31,4)·256^-4 ≈ 7e-6, i.e. ~1e-5 — the figure
+/// [`is_clean_ts`]'s own doc below states. 1-in-4-billion is the probability for
+/// four SPECIFIC packets and overstates the margin by ~4000x; at
+/// `KEY_PROOF_PACKETS = 3` the per-unit rate is ≈ C(31,3)·256^-3 ≈ 2.6e-4, so do
+/// NOT lower it on the strength of slack that is not there. It is an ABSOLUTE proof floor,
 /// NOT a proportion — a unit the key opened but whose content is bad-encoded
 /// (many non-conforming packets) is proven by ANY four good packets, not rejected
 /// for the bad ones.
@@ -172,8 +178,16 @@ pub fn is_clean(unit: &[u8], format: crate::disc::ContentFormat) -> bool {
 /// Structural "does this unit carry enough valid MPEG-TS to prove a key opened
 /// it?" — the Transport-Stream arm of [`is_clean`]. It is
 /// NOT a decryption verdict: [`decrypt_unit`] applies a key (that is
-/// "decrypt"); whether the plaintext is clean TS is this SEPARATE question. The
+/// "decrypt"); whether the plaintext is clean TS is this SEPARATE question.
-/// mux never calls this — TS validity is a muxer concern, never a decrypt result.
+///
+/// The mux is its PRINCIPAL consumer for `BdTs` discs, reaching it through
+/// [`is_clean`]: `mux::resolve`'s multi-CPS `pick` closure selects a unit key by
+/// it, `probe_index_phase` reports each FMTS index's interleave parity by it, and
+/// `decrypt::decrypt_sectors_mapped` uses it as the forensic-range verify net.
+/// (The doc used to say "the mux never calls this", which invited a maintainer to
+/// tighten or loosen the proof rule below believing only whole-disc read
+/// verification was affected — while it in fact changes which unit key a
+/// multi-CPS disc muxes with and which phase an FMTS index is muxed at.)
 ///
 /// Rule — evidence is ABSOLUTE, scaled to the packets that exist. Over the
 /// ENCRYPTED packets (skip packet 0: its `0x47` sits in the clear 16-byte seed, so
@@ -381,13 +395,20 @@ pub fn encrypt_unit(unit: &mut [u8], unit_key: &[u8; 16]) -> bool {
 /// Remove bus encryption from an aligned unit (AACS 2.0 / UHD).
 /// Bus encryption uses read_data_key, decrypting bytes 16..2048 of each 2048-byte sector.
 pub(crate) fn decrypt_bus(unit: &mut [u8], read_data_key: &[u8; 16]) {
+    // Expand the key schedule ONCE for the whole unit. `read_data_key` is
+    // loop-invariant here (and constant for the entire disc), but calling
+    // `aes_cbc_decrypt` per sector rebuilt the AES-128 schedule per sector — three
+    // expansions per 6144-byte aligned unit, i.e. ~29 million redundant expansions
+    // over a 90 GB read on a stock drive, on the per-unit decrypt hot path.
+    // Measured by `decrypt_bus_expands_the_read_data_key_once_per_unit`.
+    let cipher = crate::aacs::crypto::new_cipher_for(read_data_key);
     for sector_start in (0..ALIGNED_UNIT_LEN).step_by(SECTOR_BYTES) {
         if sector_start + SECTOR_BYTES > unit.len() {
             break;
         }
         // First 16 bytes of each sector are plaintext
-        aes_cbc_decrypt(
+        crate::aacs::crypto::cbc_decrypt_blocks(
-            read_data_key,
+            &cipher,
             &mut unit[sector_start + 16..sector_start + SECTOR_BYTES],
         );
     }
@@ -1182,21 +1203,100 @@ mod tests {
         assert_eq!(aes_ecb_decrypt(&key, &expected), pt);
     }
 
+    // ── decrypt_bus: one key schedule per unit, not one per sector ─────────
+
+    /// MEASURED, not reasoned: `decrypt_bus` called `aes_cbc_decrypt` once per
+    /// 2048-byte sector, and each call built its own AES-128 key schedule, so a
+    /// 6144-byte aligned unit performed THREE key expansions under the same
+    /// loop-invariant `read_data_key`. On a 90 GB UHD read on a stock (non-
+    /// LibreDrive) drive — ~14.6 million aligned units — that is ~29 million
+    /// redundant expansions on the per-unit decrypt hot path, for a key that is
+    /// constant for the whole disc. The counter is incremented inside
+    /// `crypto::new_cipher`, the single construction site.
+    #[test]
+    fn decrypt_bus_expands_the_read_data_key_once_per_unit() {
+        use crate::aacs::crypto::KEY_EXPANSIONS;
+        let mut unit = clear_unit();
+        let rdk = [0x4Eu8; 16];
+        KEY_EXPANSIONS.with(|c| c.set(0));
+        decrypt_bus(&mut unit, &rdk);
+        let n = KEY_EXPANSIONS.with(|c| c.get());
+        assert_eq!(
+            n, 1,
+            "one aligned unit under one read_data_key must expand the schedule \
+             exactly once, not once per 2048-byte sector"
+        );
+    }
+
+    /// The single-expansion refactor must be byte-identical: bus encryption
+    /// ([C] §4.2) covers bytes 16..2048 of every 2048-byte sector, so a
+    /// three-sector aligned unit round-trips through the forward direction
+    /// sector by sector and `decrypt_bus` must recover it exactly.
+    #[test]
+    fn decrypt_bus_roundtrips_every_sector_region() {
+        let rdk = [0x91u8; 16];
+        let original = clear_unit();
+        let mut unit = original.clone();
+        // Forward direction, region by region — the inverse of decrypt_bus.
+        for start in (0..ALIGNED_UNIT_LEN).step_by(SECTOR_BYTES) {
+            crate::aacs::crypto::aes_cbc_encrypt(&rdk, &mut unit[start + 16..start + SECTOR_BYTES]);
+        }
+        assert_ne!(
+            &unit[16..64],
+            &original[16..64],
+            "the forward direction must have changed the bytes"
+        );
+        decrypt_bus(&mut unit, &rdk);
+        assert_eq!(
+            unit.as_slice(),
+            original.as_slice(),
+            "decrypt_bus must invert the per-sector bus encryption exactly"
+        );
+    }
+
     // ── CBC decrypt: first-block uses fixed AACS IV ────────────────────────
 
+    /// The published `iv0` bytes, INDEPENDENT of the production constant.
+    ///
+    /// [C] §2.1.2 fixes one default CBC IV for every AACS AES-CBC operation.
+    /// Both IV tests below used to compute their expected value from
+    /// `crypto::AACS_IV` itself, so the constant was asserted against itself and
+    /// NOTHING in the suite pinned its bytes: swapping `AACS_IV` for `[0u8; 16]`
+    /// left both tests passing (one builds its ciphertext with the same value and
+    /// the other cancels the change in a triple XOR) while every real AACS disc
+    /// decrypted to noise — block 0 of every 6128-byte aligned unit and of every
+    /// bus-encrypted sector XORed with the wrong IV. This literal is the
+    /// independent witness the tests assert against.
+    const IV0_PUBLISHED: [u8; 16] = [
+        0x0B, 0xA0, 0xF8, 0xDD, 0xFE, 0xA6, 0x1F, 0xB3, 0xD8, 0xDF, 0x9F, 0x56, 0x6A, 0x05, 0x0F,
+        0x78,
+    ];
+
+    /// Pins the fixed AACS CBC IV ([C] §2.1.2 `iv0`) against a literal, so a
+    /// change to `crypto::AACS_IV` fails HERE rather than silently shipping.
+    #[test]
+    fn aacs_iv_matches_published_iv0() {
+        assert_eq!(
+            AACS_IV, IV0_PUBLISHED,
+            "the fixed AACS CBC IV must be the published iv0"
+        );
+    }
+
     #[test]
     fn cbc_decrypt_first_block_xors_aacs_iv() {
         // CBC: P[0] = AES-D(K, C[0]) XOR IV, and the IV is the fixed AACS
-        // constant (not zero). Encrypt a single block forward with IV, then
+        // constant (not zero). Encrypt a single block forward with the PUBLISHED
-        // confirm aes_cbc_decrypt recovers it — proving the IV used on block
+        // iv0 literal, then confirm aes_cbc_decrypt recovers it — proving the IV
-        // 0 is exactly AACS_IV. A mutation that swaps AACS_IV for [0u8;16]
+        // the production code uses on block 0 is exactly that value. Building the
-        // makes the recovered block wrong.
+        // fixture from `IV0_PUBLISHED` rather than from `AACS_IV` is what makes
+        // the claim real: a mutation that swaps AACS_IV for [0u8;16] now makes the
+        // recovered block wrong.
         let key = [0x24u8; 16];
         let plain = [0x5Au8; 16];
         // Forward CBC for one block: C = AES-E(K, P XOR IV).
         let mut x = plain;
         for j in 0..16 {
-            x[j] ^= AACS_IV[j];
+            x[j] ^= IV0_PUBLISHED[j];
         }
         let ct = aes_ecb_encrypt(&key, &x);
         let mut buf = ct;
@@ -1225,10 +1325,13 @@ mod tests {
         //   * Blocks 1..=3 are independent of the IV — they MUST equal the NIST
         //     plaintext byte-for-byte (P[i] = AES-D(K, C[i]) XOR C[i-1]). This
         //     pins the real reverse-order CBC chaining against a published KAT.
-        //   * Block 0 = AES-D(K, C[0]) XOR AACS_IV = NIST_PT[0] XOR NIST_IV
+        //   * Block 0 = AES-D(K, C[0]) XOR iv0 = NIST_PT[0] XOR NIST_IV XOR iv0 —
-        //     XOR AACS_IV — the documented IV substitution. Asserting this exact
+        //     the documented IV substitution. Asserting this exact relation pins
-        //     relation pins both the AES decrypt of C[0] AND that block 0 uses
+        //     both the AES decrypt of C[0] AND that block 0 uses iv0. The expected
-        //     AACS_IV (a swap to [0u8;16] or a chaining bug fails it).
+        //     value is built from the `IV0_PUBLISHED` literal, NOT from
+        //     `crypto::AACS_IV`: computing it from the production constant made
+        //     the change cancel out of the triple XOR, so a swap to [0u8;16] still
+        //     passed. It now fails.
         let key = [
             0x2B, 0x7E, 0x15, 0x16, 0x28, 0xAE, 0xD2, 0xA6, 0xAB, 0xF7, 0x15, 0x88, 0x09, 0xCF,
             0x4F, 0x3C,
@@ -1268,7 +1371,7 @@ mod tests {
         // Block 0: NIST_PT[0] XOR NIST_IV XOR AACS_IV (the fixed-IV substitution).
         let mut expected_block0 = [0u8; 16];
         for i in 0..16 {
-            expected_block0[i] = nist_plaintext[i] ^ nist_iv[i] ^ AACS_IV[i];
+            expected_block0[i] = nist_plaintext[i] ^ nist_iv[i] ^ IV0_PUBLISHED[i];
         }
         assert_eq!(
             &buf[0..16],

src/aacs/crypto.rs

@@ -15,6 +15,33 @@ pub(crate) const AACS_IV: [u8; 16] = [
     0x0B, 0xA0, 0xF8, 0xDD, 0xFE, 0xA6, 0x1F, 0xB3, 0xD8, 0xDF, 0x9F, 0x56, 0x6A, 0x05, 0x0F, 0x78,
 ];
 
+// Per-thread count of AES-128 key schedules built through `new_cipher`.
+// Test-only instrumentation: an AES-128 key expansion is 10 round-key
+// derivations, and the CBC helpers here run on the per-aligned-unit decrypt hot
+// path of a whole disc read, so "how many times was the schedule built for one
+// loop-invariant key" is a property worth asserting rather than reasoning about.
+// THREAD-LOCAL, not a global atomic: `cargo test` runs tests concurrently, so a
+// shared counter would see every other test's expansions. See
+// `content::tests::decrypt_bus_expands_the_read_data_key_once_per_unit`.
+#[cfg(test)]
+thread_local! {
+    pub(crate) static KEY_EXPANSIONS: std::cell::Cell<usize> = const { std::cell::Cell::new(0) };
+}
+
+/// Build an AES-128 key schedule for a caller that will drive
+/// [`cbc_decrypt_blocks`] over several regions under one key.
+pub(crate) fn new_cipher_for(key: &[u8; 16]) -> Aes128 {
+    new_cipher(key)
+}
+
+/// Build an AES-128 key schedule. The single construction site for the CBC
+/// helpers, so [`KEY_EXPANSIONS`] can count them under test.
+fn new_cipher(key: &[u8; 16]) -> Aes128 {
+    #[cfg(test)]
+    KEY_EXPANSIONS.with(|c| c.set(c.get() + 1));
+    Aes128::new(GenericArray::from_slice(key))
+}
+
 /// AES-128-ECB encrypt a single 16-byte block. [C] §2.1.1 (`AES-128E`).
 pub(crate) fn aes_ecb_encrypt(key: &[u8; 16], data: &[u8; 16]) -> [u8; 16] {
     let cipher = Aes128::new(GenericArray::from_slice(key));
@@ -35,13 +62,12 @@ pub(crate) fn aes_ecb_decrypt(key: &[u8; 16], data: &[u8; 16]) -> [u8; 16] {
     out
 }
 
-/// AES-128-CBC decrypt in-place with the fixed AACS IV. [C] §2.1.2 (`AES-128CBCD`).
+/// AES-128-CBC ENCRYPT in place under the fixed [`AACS_IV`] — the forward
+/// direction of [`aes_cbc_decrypt`], and its exact inverse. [C] §2.1.2
+/// (`AES-128CBCE`).
 ///
-/// Precondition: `data.len()` is a multiple of 16. Any trailing partial
+/// Precondition: `data.len()` is a multiple of 16; the assert
-/// block is silently ignored; all callers pass aligned regions (6128 and
+/// documents/enforces that contract.
-/// 2032 bytes), and the assert documents/enforces that contract.
-/// AES-128-CBC encrypt in place under the fixed [`AACS_IV`] — the forward
-/// direction of [`aes_cbc_decrypt`], and its exact inverse.
 ///
 /// Constructs the cipher ONCE for the whole slice. Driving this from the
 /// single-block [`aes_ecb_encrypt`] instead rebuilds the AES key schedule per
@@ -52,7 +78,7 @@ pub(crate) fn aes_cbc_encrypt(key: &[u8; 16], data: &mut [u8]) {
         data.len().is_multiple_of(16),
         "aes_cbc_encrypt requires a block-aligned slice"
     );
-    let cipher = Aes128::new(GenericArray::from_slice(key));
+    let cipher = new_cipher(key);
     let num_blocks = data.len() / 16;
     let mut prev = AACS_IV;
     // Forward order: each block is XORed with the PRECEDING ciphertext block.
@@ -69,12 +95,40 @@ pub(crate) fn aes_cbc_encrypt(key: &[u8; 16], data: &mut [u8]) {
     }
 }
 
+/// AES-128-CBC DECRYPT in-place with the fixed AACS IV. [C] §2.1.2
+/// (`AES-128CBCD`).
+///
+/// Precondition: `data.len()` is a multiple of 16. Any trailing partial
+/// block is silently ignored; all callers pass aligned regions (6128 and
+/// 2032 bytes), and the assert documents/enforces that contract.
+///
+/// (This doc block was orphaned onto `aes_cbc_encrypt` above when that function
+/// was inserted directly after it with no separating blank line, so rustdoc
+/// rendered the crate's only forward-direction AACS primitive as "decrypt" and
+/// cited the spec's DECRYPT clause for it, while this function had no doc at
+/// all. `encrypt_unit_is_the_exact_inverse_of_decrypt_unit` in `content.rs` pins
+/// the directions behaviourally so a maintainer 'fixing' the contradiction by
+/// swapping the two bodies fails the suite instead of shipping a second
+/// decryptor behind an already-set encrypted flag.)
 pub(crate) fn aes_cbc_decrypt(key: &[u8; 16], data: &mut [u8]) {
     debug_assert!(
         data.len().is_multiple_of(16),
         "aes_cbc_decrypt requires a block-aligned slice"
     );
-    let cipher = Aes128::new(GenericArray::from_slice(key));
+    cbc_decrypt_blocks(&new_cipher(key), data);
+}
+
+/// AES-128-CBC decrypt in place under the fixed [`AACS_IV`] with an ALREADY
+/// EXPANDED key schedule.
+///
+/// Split out of [`aes_cbc_decrypt`] so a caller that decrypts several regions
+/// under one loop-invariant key expands the schedule once. `decrypt_bus`
+/// ([`super::content::decrypt_bus`]) is that caller: bus encryption
+/// ([C] §4.2 / the AACS 2.0 Read Data Key) covers bytes 16..2048 of EVERY
+/// 2048-byte sector, so a 6144-byte aligned unit is three regions under one
+/// `read_data_key` — three key schedules where one suffices, on the per-unit
+/// decrypt hot path of a whole 90 GB read.
+pub(crate) fn cbc_decrypt_blocks(cipher: &Aes128, data: &mut [u8]) {
     let num_blocks = data.len() / 16;
     // Process blocks in reverse to avoid clobbering ciphertext needed for XOR
     for i in (0..num_blocks).rev() {

src/decrypt.rs

@@ -1240,6 +1240,62 @@ mod tests {
         }
     }
 
+    /// The mapped descramble indexes the committed key pool POSITIONALLY
+    /// (`unit_keys[key_idx].1`), so the ORDER of the `Vec<UnitKey>` a
+    /// `KeySource` returns is load-bearing — it is NOT "cosmetic, the decrypt path
+    /// strips it and tries every key", as `keysource::resolve_and_apply_traced`'s
+    /// doc used to claim. Trial-decrypt was deliberately deleted; nothing here
+    /// searches the pool. Reordering the same two keys therefore sends each range
+    /// to the WRONG key: the range that decrypted clean now fails the correct-phase
+    /// `is_clean` net loudly (or, off a forensic phase, would decrypt a whole span
+    /// under a neighbour's key). Pins the corrected doc.
+    #[test]
+    fn mapped_key_selection_is_positional_so_pool_order_matters() {
+        use crate::disc::ContentFormat;
+        let key_a = [0xAAu8; 16];
+        let key_b = [0xBBu8; 16];
+        let ul = aacs::content::ALIGNED_UNIT_LEN;
+        let usz = (ul / 2048) as u32;
+        // Unit 0 encrypted under key_a, unit 1 under key_b.
+        let build = || {
+            let mut buf = vec![0u8; 2 * ul];
+            for (i, k) in [key_a, key_b].iter().enumerate() {
+                let mut u = clear_ts_unit();
+                aacs_encrypt_unit_for_test(&mut u, k);
+                buf[i * ul..(i + 1) * ul].copy_from_slice(&u);
+            }
+            buf
+        };
+        // Map: unit 0 → pool position 0, unit 1 → pool position 1.
+        let map = AacsKeyMap::from_ranges_phased(vec![
+            (0, usz, 0, Phase::Even),
+            (usz, 2 * usz, 1, Phase::Even),
+        ]);
+        // Pool in CPS-unit order: each range gets its own key, both come clean.
+        let mut buf = build();
+        let keys = DecryptKeys::Aacs {
+            unit_keys: vec![(0, key_a), (1, key_b)],
+            read_data_key: None,
+            format: ContentFormat::BdTs,
+        };
+        decrypt_sectors_mapped(&mut buf, &keys, 0, &map)
+            .expect("pool in CPS-unit order decrypts clean");
+        // SAME keys, SAME CPS-unit numbers, swapped POSITIONS. If the number were
+        // what mattered (or if the path searched the pool) this would be
+        // equivalent; positional indexing makes it decrypt both units wrong.
+        let mut buf = build();
+        let swapped = DecryptKeys::Aacs {
+            unit_keys: vec![(1, key_b), (0, key_a)],
+            read_data_key: None,
+            format: ContentFormat::BdTs,
+        };
+        assert!(
+            decrypt_sectors_mapped(&mut buf, &swapped, 0, &map).is_err(),
+            "a reordered pool must fail loud — key selection is positional, so the \
+             ORDER a KeySource returns its keys in is part of the contract"
+        );
+    }
+
     /// The correct-phase safety `is_clean` fires loud: an even unit whose mapped
     /// key is wrong does NOT come clean → `DecryptFailed` (not silent corruption).
     #[test]

src/drive/mod.rs

@@ -781,9 +781,34 @@ impl Drive {
         // each a self-contained READ(10) with the same validation. Any
         // chunk error reports that chunk's LBA (more precise than the whole
         // request's base LBA).
+        let count = count as u32;
+        // Check the caller's buffer ONCE, up front. The chunk loop slices `buf`
+        // by `count * 2048`; without this an undersized buffer PANICKED ('range
+        // end index out of range') out of the public `read`/`read_fua`, while the
+        // single-chunk path above tolerates the same undersized buffer and returns
+        // `Err(DiscRead)` from `checked_exec`. Behaviour on an undersized buffer
+        // must not depend on the transport's transfer limit.
+        if buf.len() < count as usize * 2048 {
+            return Err(Error::DiscRead {
+                sector: lba as u64,
+                status: None,
+                sense: None,
+            });
+        }
+        // The whole range must be addressable: SBC-3 READ(10) carries a 32-bit
+        // LOGICAL BLOCK ADDRESS, so a request whose last chunk crosses `u32::MAX`
+        // has no valid CDB. `lba + done` below was unchecked — a debug panic out
+        // of the public API, and in release a wrap to a low LBA that was read and
+        // returned as if it were the requested one.
+        if lba.checked_add(count.saturating_sub(1)).is_none() {
+            return Err(Error::DiscRead {
+                sector: lba as u64,
+                status: None,
+                sense: None,
+            });
+        }
         let mut done: u32 = 0;
         let mut total: usize = 0;
-        let count = count as u32;
         while done < count {
             let chunk = (count - done).min(max_sectors);
             let cur_lba = lba + done;
@@ -1935,6 +1960,54 @@ mod command_tests {
         assert_eq!(*reads.lock().unwrap(), vec![(0, 3)], "single CDB, no split");
     }
 
+    /// The multi-chunk path slices the caller's buffer by `count * 2048` with no
+    /// length check, so an undersized `buf` PANICKED ('range end index out of
+    /// range') out of the public `Drive::read` / `Drive::read_fua` — while the
+    /// single-chunk path (`read_one` → `checked_exec`) tolerates the same
+    /// undersized buffer and returns `Err(DiscRead)`. The public API's behaviour
+    /// on an undersized buffer must not depend on the transport's transfer limit.
+    #[test]
+    fn undersized_buffer_multi_chunk_errors_not_panics() {
+        let ChunkingHarness {
+            drive: mut d,
+            reads: _reads,
+        } = chunking(4 * 2048, None);
+        // count (10) > max_sectors (4) → the chunk loop; buf holds only 1 sector.
+        let mut buf = vec![0u8; 2048];
+        assert!(
+            matches!(d.read(0, 10, &mut buf, false), Err(Error::DiscRead { .. })),
+            "an undersized buffer must error, not panic"
+        );
+        // The single-chunk path with the SAME undersized buffer already errored;
+        // the two paths must now agree.
+        let mut buf = vec![0u8; 2048];
+        assert!(
+            matches!(d.read(0, 3, &mut buf, false), Err(Error::DiscRead { .. })),
+            "single-chunk path errors on an undersized buffer (unchanged)"
+        );
+    }
+
+    /// `Drive::read`'s chunk loop advanced the per-chunk LBA with an unchecked
+    /// `lba + done`. SBC-3 READ(10) `LOGICAL BLOCK ADDRESS` is a 32-bit field, so
+    /// a request whose last chunk crosses `u32::MAX` overflowed: debug panic out
+    /// of the public API, release wrap to a low LBA silently read instead.
+    #[test]
+    fn chunk_lba_near_u32_max_errors_not_overflows() {
+        let ChunkingHarness {
+            drive: mut d,
+            reads: _reads,
+        } = chunking(4 * 2048, None);
+        let mut buf = vec![0u8; 10 * 2048];
+        // 0xFFFF_FFFE + 4 overflows on the second chunk.
+        assert!(
+            matches!(
+                d.read(0xFFFF_FFFE, 10, &mut buf, false),
+                Err(Error::DiscRead { .. })
+            ),
+            "an LBA range past u32::MAX must error, not overflow"
+        );
+    }
+
     // ── find_drive media-preference selection policy ────────────────
 
     /// Build a fake drive whose GET EVENT STATUS reply reports the given

src/io/pipeline.rs

@@ -125,7 +125,10 @@ fn consumer_panicked(payload: Box<dyn std::any::Any + Send>) -> Error {
 /// caller reporting the rip as interrupted while a fully finalised MKV (Cues
 /// written, Segment size patched) landed on disk, indistinguishable from a
 /// complete one. The two transitions are therefore a single compare-exchange each,
-/// out of [`ST_RUNNING`]: whoever wins decides, and the loser observes the winner.
+/// out of [`state::RUNNING`]: whoever wins decides, and the loser observes the
+/// winner. (`ST_RUNNING` does not exist anywhere in the crate — the constants are
+/// `state::RUNNING` / `state::ABANDONED` / `state::CLOSING` below, and both
+/// compare-exchange sites that must stay in step with this argument name them.)
 mod state {
     /// Consumer is running; neither side has committed yet.
     pub const RUNNING: u8 = 0;

src/io/writeback_file/macos.rs

@@ -112,8 +112,26 @@ pub(super) fn durable_sync(file: &File) -> io::Result<()> {
             );
             Ok(())
         }
-        Err(crate::io::bounded::BoundedError::Halted) => Ok(()),
+        // Both arms below used to map to `Ok(())` with NO diagnostic at all, while
-        Err(crate::io::bounded::BoundedError::WorkerLost) => Ok(()),
+        // the Linux sibling logs the identical failures (writeback_file/linux.rs).
+        // A lost F_FULLFSYNC worker at the end of a UHD mux therefore reported
+        // `completed = true` with an empty log, leaving an operator investigating a
+        // truncated/corrupt output file after a power loss no record that the final
+        // fsync never ran — on Linux the same failure is at error level.
+        Err(crate::io::bounded::BoundedError::Halted) => {
+            tracing::warn!(
+                target: "mux",
+                "WritebackFile::sync_all F_FULLFSYNC skipped (halt requested); data not durably flushed, kernel will flush on close"
+            );
+            Ok(())
+        }
+        Err(crate::io::bounded::BoundedError::WorkerLost) => {
+            tracing::error!(
+                target: "mux",
+                "WritebackFile::sync_all F_FULLFSYNC worker lost before completion; data not durably flushed, kernel will flush on close"
+            );
+            Ok(())
+        }
     }
 }
 

src/keysource.rs

@@ -53,9 +53,30 @@ pub const MIN_SAMPLE_UNITS: usize = 8;
 /// units it yields); the *requested* count is a caller-side compile-time constant that
 /// callers pin to `MIN_SAMPLE_UNITS` (see e.g. autorip's `SAMPLE_UNITS`). Together the
 /// two make under-sampling unrepresentable at the request boundary.
-#[derive(Debug, Clone)]
+///
+/// The wrapped samples are on-disc AACS ciphertext — the same bytes the sibling
+/// [`DiscInputs::samples`] redacts as key MATERIAL — so [`Debug`] is hand-written
+/// and redacting; see the impl below.
+#[derive(Clone)]
 pub struct DecodeSampleSet(Vec<Vec<u8>>);
 
+impl std::fmt::Debug for DecodeSampleSet {
+    /// Prints the SHAPE only. A derived `Debug` dumped every wrapped sample
+    /// verbatim: a `DecodeSampleSet` carries at least [`MIN_SAMPLE_UNITS`]
+    /// 6144-byte aligned units (≥ 49 KiB, in practice multi-MB) of AACS
+    /// ciphertext plus each unit's clear 16-byte derivation seed, so one
+    /// `tracing::debug!("{set:?}")` on a failed `/decode` request — or an
+    /// `assert_eq!` whose panic message formats it — wrote all of it to the log
+    /// that gets attached to a bug report. Same policy and same shape as
+    /// [`DiscInputs`]'s impl below.
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("DecodeSampleSet")
+            .field("units", &"<redacted>")
+            .field("units_len", &self.0.len())
+            .finish()
+    }
+}
+
 impl DecodeSampleSet {
     /// Wrap `units` iff it carries at least [`MIN_SAMPLE_UNITS`] samples; `None`
     /// otherwise (the caller then skips the online source rather than sending an
@@ -334,8 +355,23 @@ pub fn resolve_and_apply(
 /// from [`crate::aacs::derive::decrypt_unit_key`]; the library's canonical CPS-unit number is
 /// `position + 1` (matching [`crate::aacs::inf::parse_unit_key_ro`]'s `(i + 1)`), so
 /// the committed `AacsState.unit_keys` is byte-identical to the library-resolved
-/// path. The number is cosmetic for descramble (the decrypt path strips it and
+/// path.
-/// tries every key) but is kept faithful to the resolver's convention.
+///
+/// The NUMBER itself is not what descramble indexes by — but the ORDER is
+/// load-bearing, so a source must return its keys in CPS-unit order. Trial
+/// decrypt-and-check was deliberately deleted (see
+/// [`crate::decrypt::AacsKeyMap`]: decryption is driven by the disc's CPS-unit /
+/// FMTS-segment structure, "never by trial-decrypt-and-check per unit"), and
+/// `decrypt_sectors_mapped` indexes the committed pool POSITIONALLY —
+/// `unit_keys[key_idx].1`, where `key_idx` is a POSITION in the Vec a source
+/// returned, recorded by `resolve_mux_key_map_cached` / `resolve_fmts_key_map`.
+/// Return the same keys in a different order and every `AacsKeyMap` points at the
+/// wrong key: the whole title decrypts under a neighbour's key, or a forensic
+/// range trips the `is_clean` net into `DecryptFailed`. (The doc used to say the
+/// number "is cosmetic for descramble (the decrypt path strips it and tries every
+/// key)", which is what the DELETED trial-decrypt path did; the only place that
+/// still tries every key is `Disc::decrypt_with`'s sample VALIDATION, which does
+/// not descramble content.)
 pub fn resolve_and_apply_traced(
     sources: &[Box<dyn KeySource>],
     inputs: &DiscInputs,
@@ -400,8 +436,15 @@ pub fn resolve_and_apply_traced(
             // boundary in `freemkv-keysources`, so a failure is reported as a
             // failure, with `Disc::aacs_error` as the channel the operator actually
             // reads. `FetchOutcome::errored` in `drive_unit_keys` /
-            // `drive_fmts_indexes` is dead for the same reason — it is the right
+            // `drive_fmts_indexes` never FIRES for the same reason — it is the
-            // contract, honoured by no source yet.
+            // right contract, honoured by no shipped source yet. It is NOT dead
+            // code: it is written at both `drive_*` sites and read by the
+            // cache-insert guard `if !keys.is_empty() || !outcome.errored`, the
+            // only thing that stops a transient source outage from being memoised
+            // permanently into the per-fingerprint key cache — pinned by
+            // `errored_empty_is_not_cached_and_retries_when_source_recovers`. Do
+            // not delete it while making a source report failures as `Err`; that
+            // is precisely when it starts to matter.
             Ok(_) | Err(_) => {
                 trace.keys.push(KeyStep {
                     who,
@@ -1396,4 +1439,28 @@ mod tests {
         assert!(dbg.contains("samples_len: 1"), "{dbg}");
         assert!(dbg.contains("TITLE_2024"), "{dbg}");
     }
+
+    /// `DecodeSampleSet` is public and wraps the SAME on-disc ciphertext the
+    /// sibling `DiscInputs` redacts, so a derived `Debug` dumped ≥ MIN_SAMPLE_UNITS
+    /// × 6144 bytes of verbatim AACS ciphertext (plus every unit's clear 16-byte
+    /// derivation seed) into any log that formatted it. Sentinel byte 0xD5 =
+    /// decimal 213, matching `aacs::types::redaction_tests` and the
+    /// `DiscInputs` test above. Mutation guard: restoring `#[derive(Debug)]`
+    /// fails this.
+    #[test]
+    fn decode_sample_set_debug_is_redacted() {
+        let set = DecodeSampleSet::new(vec![vec![0xD5; 6144]; MIN_SAMPLE_UNITS])
+            .expect("MIN_SAMPLE_UNITS units is a valid set");
+        let dbg = format!("{set:?}");
+        assert!(
+            !dbg.contains("213"),
+            "DecodeSampleSet Debug leaked ciphertext (decimal 213): {dbg}"
+        );
+        assert!(
+            dbg.contains("redacted"),
+            "DecodeSampleSet Debug missing redaction marker: {dbg}"
+        );
+        // Non-secret shape stays printable for diagnostics.
+        assert!(dbg.contains("units_len: 8"), "{dbg}");
+    }
 }

src/mux/au_assembly.rs

@@ -126,6 +126,12 @@ pub(crate) struct AuAssembler {
     /// so a long run of junk with no start code (hostile/corrupt input) costs
     /// O(bytes) total, not O(buffer) per push. Reset when `buf[0]` moves.
     opener_pos: usize,
+    /// Test-only: how many times `take_front` fell back to the COPY path. The
+    /// handover is the whole point of `take_front`, so "did it actually fire" is a
+    /// property to MEASURE, not to reason about. See
+    /// `handover_survives_a_large_au_instead_of_copying_every_later_one`.
+    #[cfg(test)]
+    copy_path_hits: usize,
 }
 
 impl AuAssembler {
@@ -155,6 +161,8 @@ impl AuAssembler {
             scan_pos: 0,
             seen_unit: false,
             opener_pos: 0,
+            #[cfg(test)]
+            copy_path_hits: 0,
         }
     }
 
@@ -172,6 +180,8 @@ impl AuAssembler {
             scan_pos: 0,
             seen_unit: false,
             opener_pos: 0,
+            #[cfg(test)]
+            copy_path_hits: 0,
         }
     }
 
@@ -356,14 +366,38 @@ impl AuAssembler {
     /// (a small AU after a multi-MB one): handing over would otherwise attach an
     /// oversized idle allocation to a small frame for as long as the frame queues
     /// downstream, trading a copy for resident memory.
+    ///
+    /// That fallback must not become permanent. `buf`'s capacity used to be a
+    /// one-way high-water mark — the replacement buffer was created with
+    /// `cap.max(tail_len)`, and the copy path's `drain` also preserves `cap` — so
+    /// once ONE large AU had been assembled, every later smaller AU satisfied
+    /// `cap > 2*end` and took the copy path forever. On a UHD HEVC title the first
+    /// IDR grows `buf` to ~4-8 MB, after which each ~200-400 KB P/B AU paid a
+    /// whole-AU allocation plus a whole-AU memcpy plus a tail memmove for ~99% of
+    /// the ~200,000 coded pictures — tens of GB of exactly the memcpy this handover
+    /// exists to remove. So the copy path now also RELEASES the high-water
+    /// capacity, which re-arms the handover for the next AU: one copy after a size
+    /// step down, not one per frame forever.
     fn take_front(&mut self, end: usize) -> Vec<u8> {
         let cap = self.buf.capacity();
+        let tail_len = self.buf.len() - end;
         if cap > end.saturating_mul(2) {
+            #[cfg(test)]
+            {
+                self.copy_path_hits += 1;
+            }
             let data = self.buf[..end].to_vec();
             self.buf.drain(..end);
+            // Shrink toward what this AU actually needed (the tail plus room for
+            // another AU of about this size). Only the short tail is copied, and it
+            // brings `cap` back under the `2*end` threshold so the next AU of this
+            // size hands over instead of copying.
+            self.buf.shrink_to(end.max(tail_len));
             return data;
         }
-        let mut tail = Vec::with_capacity(cap.max(self.buf.len() - end));
+        // Replacement buffer: enough for the tail plus room to accumulate the next
+        // AU of about this size. NOT `cap`, which would re-pin the high-water mark.
+        let mut tail = Vec::with_capacity(end.max(tail_len));
         tail.extend_from_slice(&self.buf[end..]);
         let mut data = std::mem::replace(&mut self.buf, tail);
         data.truncate(end);
@@ -915,11 +949,56 @@ mod tests {
             before,
             "the emitted AU must own the buffer's allocation (no whole-frame copy)"
         );
-        assert_eq!(
+        // The replacement buffer keeps room for another AU of about this size, so
+        // the next AU does not re-grow — but it is NOT pinned to the OLD capacity,
+        // which would make `buf` a permanent high-water mark and send every later
+        // smaller AU down the copy path (see
+        // `handover_survives_a_large_au_instead_of_copying_every_later_one`).
+        assert!(
+            a.buf.capacity() >= au1.len(),
+            "replacement buffer must fit another AU of this size: {} < {}",
             a.buf.capacity(),
-            cap_before,
+            au1.len()
-            "the replacement buffer keeps the capacity, so the next AU does not re-grow"
+        );
+        assert!(
+            a.buf.capacity() <= cap_before,
+            "replacement buffer must never EXCEED the old capacity"
         );
         assert_eq!(a.buf.len(), 4, "the buffer holds only AU2's delimiter tail");
     }
+
+    /// MEASURED: `take_front`'s copy fallback must not become permanent.
+    ///
+    /// `buf`'s capacity used to be a one-way high-water mark, and the copy path's
+    /// `drain` preserves it, so after ONE large AU every later smaller AU satisfied
+    /// `cap > 2*end` and copied forever. On a UHD HEVC title the first IDR grows
+    /// `buf` to multiple MB, after which ~99% of the ~200,000 coded pictures each
+    /// paid a whole-AU allocation + whole-AU memcpy + tail memmove — tens of GB of
+    /// exactly the copy the handover exists to remove. Counted at the copy path
+    /// itself: one copy is expected right after the size step down; a per-frame
+    /// copy is the bug.
+    #[test]
+    fn handover_survives_a_large_au_instead_of_copying_every_later_one() {
+        let mut a = AuAssembler::for_codec(Codec::H264);
+        // Production shape: BD-TS aligns one access unit per PES, so each `push`
+        // carries about one AU and the buffer holds ~one AU at a time. One large AU
+        // (the IDR) followed by a run of much smaller ones (P/B frames). Each AU is
+        // pushed with the NEXT AU's opener so the previous one closes.
+        const SMALL: usize = 64 * 1024;
+        let mut pending = au(0x11, 2 * 1024 * 1024);
+        for i in 0..20u8 {
+            let next = au(0x30 + i, SMALL);
+            // Append the next AU's 4-byte opener to close `pending`, push, and
+            // carry the rest of `next` forward.
+            pending.extend_from_slice(&next[..4]);
+            a.push(&pending, Some(1), None, None, false);
+            pending = next[4..].to_vec();
+        }
+        let hits = a.copy_path_hits;
+        assert!(
+            hits <= 2,
+            "the copy fallback must re-arm the handover, not fire for every AU \
+             after a large one: {hits} copies over 20 access units"
+        );
+    }
 }

src/mux/codec/hevc.rs

@@ -138,13 +138,28 @@ pub struct HevcParser {
     // is consumed (cleared) by that first CRA so only ONE CRA per boundary is
     // touched — never a mid-stream CRA, never an IDR, never a non-CRA NAL.
     //
-    // SAFETY: defaults to `false` and is ONLY ever set through
+    // ARMED BY TWO PATHS — do not read this flag as caller-driven only:
-    // `mark_clip_boundary`, which the caller invokes ONLY for a non-seamless
+    //
+    //  1. `mark_clip_boundary`, which a caller invokes only for a non-seamless
     //     (0x05/0x06) join. connection_condition 0x01 is the first-item/seamless
-    // case and must NOT trigger this flag. A stream with no boundary marker
+    //     case and must NOT trigger this flag. In practice NO caller wires this
-    // (single-clip title, or seamless-joined 0x01 UHD/BD) never has this set,
+    //     up: the mpls connection_condition is not plumbed through the threaded
-    // so the rewrite branch is never reached and output is byte-identical to a
+    //     mux pipeline (see the note at the auto-detect site in `parse`).
-    // parser without this field.
+    //  2. The in-parser PTS-backstep AUTO-DETECTION in `parse` — a backward PES
+    //     PTS step beyond `BACKSTEP_TICKS` sets it with no caller involvement.
+    //     This is the path that actually fires in production, and it is the one
+    //     the CRA→BLA rewrite exists for.
+    //
+    // So the rewrite branch is NOT dead, and output is NOT byte-identical to a
+    // parser without this field: any stream whose PES PTS steps backward by more
+    // than `BACKSTEP_TICKS` — including a damaged/rewritten PTS field on an
+    // untrusted disc — has its next CRA_NUT rewritten to BLA_W_LP, which makes a
+    // decoder discard that CRA's valid RASL leading pictures. That false-arming
+    // risk is held down by the `PTS_WRAP_PERIOD` unwrapping and the high-water
+    // watermark, not by the flag being unreachable: REMOVING either guard on the
+    // strength of "only `mark_clip_boundary` sets this" is a live corruption bug.
+    // Pinned by `cra_at_auto_detected_pts_backstep_rewritten_to_bla` and
+    // `cra_after_33bit_pts_wrap_not_rewritten`.
     pending_clip_boundary: bool,
     // Highest PES PTS seen on this video stream so far, on a MONOTONIC 64-bit
     // timeline (raw 33-bit PTS unwrapped across 2^33 wraparounds — see
@@ -216,6 +231,24 @@ const _: () = assert!(
     "HEVC BACKSTEP_TICKS must mirror mux::timeline::DISCONTINUITY_BACKSTEP_NS"
 );
 
+// Bytes reserved at the front of every assembled access unit so the keyframe
+// parameter-set re-assert can be spliced in without reallocating. A VPS + SPS +
+// PPS re-assert is a few hundred bytes (each a 4-byte length prefix plus a NAL
+// that is tens to low hundreds of bytes on real BD/UHD streams); 1 KiB covers it
+// with margin, and costs 1 KiB of slack per in-flight frame. If a stream's
+// parameter sets ever exceed this the splice still produces correct output — it
+// just reallocates once, exactly as it always did.
+const PARAM_REASSERT_HEADROOM: usize = 1024;
+
+// Per-thread count of keyframe re-asserts that had to reallocate the frame buffer.
+// Test-only instrumentation: the whole point of `PARAM_REASSERT_HEADROOM` is that
+// the splice is in-place, so that is MEASURED rather than reasoned about. See
+// `keyframe_param_reassert_does_not_reallocate_the_frame`.
+#[cfg(test)]
+thread_local! {
+    static PARAM_REASSERT_REALLOCS: std::cell::Cell<usize> = const { std::cell::Cell::new(0) };
+}
+
 // The 33-bit 90 kHz PES PTS counter wraps at 2^33 ticks (~26.5 h). When the raw
 // PTS steps backward by approximately a full period — i.e. it landed just past
 // the wrap — it is a counter wraparound, NOT a clip reset: unwrap it (add 2^33)
@@ -302,6 +335,16 @@ impl HevcParser {
     /// Unknown payload types are skipped by their size so a later HDR10 message
     /// in the same NAL is still reached.
     fn scan_sei(&mut self, nal: &[u8]) {
+        // Both HDR10 messages are per-stream constants and STICKY (first seen
+        // wins), so once both are captured every remaining match arm below
+        // declines and the whole scan is a guaranteed no-op. Return BEFORE
+        // `strip_emulation_prevention`, which allocates and byte-copies the entire
+        // SEI RBSP: an HDR10 UHD stream carries a prefix SEI per access unit, so
+        // without this the other ~200,000 access units of a title each paid one
+        // allocation and one copy for a result that is discarded.
+        if self.sei_mastering.is_some() && self.sei_content_light.is_some() {
+            return;
+        }
         let Some(raw) = nal.get(2..) else {
             return;
         };
@@ -532,7 +575,11 @@ impl CodecParser for HevcParser {
         // Pre-size: output is ~input bytes with a few 4-byte length
         // prefixes added. UHD frames are 150-300 KB; the unsized Vec
         // growth chain otherwise reallocs 5-7× per frame.
-        let mut frame_data = Vec::with_capacity(data.len() + 64);
+        //
+        // Plus `PARAM_REASSERT_HEADROOM` so the keyframe parameter-set re-assert
+        // below can be spliced in FRONT of the frame without reallocating. See
+        // that site.
+        let mut frame_data = Vec::with_capacity(data.len() + 64 + PARAM_REASSERT_HEADROOM);
 
         // Single-pass NAL scan: extract params, detect keyframes, build length-prefixed output
         let mut pos = 0;
@@ -670,13 +717,34 @@ impl CodecParser for HevcParser {
         // when active == codecPrivate) so each keyframe is self-contained and a
         // decoder that dropped the set (CRA reset / SPS event) self-heals.
         if keyframe {
-            let mut prefix = Vec::new();
+            let mut prefix = Vec::with_capacity(PARAM_REASSERT_HEADROOM);
             reassert_active(&mut prefix, &self.cur_vps, emitted_vps);
             reassert_active(&mut prefix, &self.cur_sps, emitted_sps);
             reassert_active(&mut prefix, &self.cur_pps, emitted_pps);
             if !prefix.is_empty() {
-                prefix.extend_from_slice(&frame_data);
+                // SPLICE the few hundred prefix bytes into the front of the
-                frame_data = prefix;
+                // already-assembled frame, in place.
+                //
+                // This used to be `prefix.extend_from_slice(&frame_data)` followed
+                // by `frame_data = prefix`: that grew `prefix` from a few hundred
+                // bytes to the FULL access-unit size (a fresh multi-MB allocation),
+                // memcpy'd the whole frame into it, and dropped the presized
+                // `frame_data` buffer — one extra whole-frame allocation plus one
+                // extra whole-frame copy per keyframe. A 2 h UHD title at 24 fps
+                // with a 1 s GOP is ~7,200 keyframes, i.e. ~7,200 multi-MB
+                // allocations and ~14-28 GB of avoidable memcpy per title.
+                //
+                // `frame_data` was reserved with `PARAM_REASSERT_HEADROOM` to spare
+                // precisely so this splice fits without reallocating; what remains
+                // is one in-place memmove inside the existing buffer. Byte-identical
+                // output either way.
+                #[cfg(test)]
+                let cap_before = frame_data.capacity();
+                frame_data.splice(0..0, prefix);
+                #[cfg(test)]
+                if frame_data.capacity() != cap_before {
+                    PARAM_REASSERT_REALLOCS.with(|c| c.set(c.get() + 1));
+                }
             }
         }
 
@@ -839,9 +907,22 @@ struct SpsChroma {
     temporal_id_nesting_flag: u8,
 }
 
+// Per-thread count of `strip_emulation_prevention` calls. Test-only
+// instrumentation: the function allocates and byte-copies a whole RBSP, and
+// `scan_sei` used to run it for every SEI NAL of every access unit, so "how many
+// copies did a stream actually cost" is worth MEASURING rather than reasoning
+// about. Thread-local, not a global atomic, because `cargo test` runs tests
+// concurrently. See `scan_sei_stops_copying_once_both_hdr10_messages_are_captured`.
+#[cfg(test)]
+thread_local! {
+    static RBSP_COPIES: std::cell::Cell<usize> = const { std::cell::Cell::new(0) };
+}
+
 /// Strip HEVC/H.264 emulation-prevention bytes (00 00 03 → 00 00) from a NAL
 /// RBSP so a bit reader sees the true coded values.
 fn strip_emulation_prevention(rbsp: &[u8]) -> Vec<u8> {
+    #[cfg(test)]
+    RBSP_COPIES.with(|c| c.set(c.get() + 1));
     let mut out = Vec::with_capacity(rbsp.len());
     let mut zeros = 0usize;
     for &b in rbsp {
@@ -1169,6 +1250,67 @@ mod tests {
         assert_eq!(h.max_pic_average_light_level, maxfall);
     }
 
+    /// MEASURED, not reasoned: an HDR10 stream carries a prefix SEI per access
+    /// unit, and `scan_sei` allocated + byte-copied the whole SEI RBSP through
+    /// `strip_emulation_prevention` on EVERY one — including after both HDR10
+    /// messages were already captured and every match arm was guaranteed to
+    /// decline. On a ~200,000-frame UHD title that is ~200,000 allocations and
+    /// copies for a discarded result. Counted at the single
+    /// `strip_emulation_prevention` site.
+    #[test]
+    fn scan_sei_stops_copying_once_both_hdr10_messages_are_captured() {
+        let pps = {
+            let mut v = vec![0x00, 0x00, 0x01];
+            v.extend_from_slice(&hevc_nal_header(NAL_PPS));
+            v.push(0xC0);
+            v
+        };
+        let idr = {
+            let mut v = vec![0x00, 0x00, 0x01];
+            v.extend_from_slice(&hevc_nal_header(19));
+            v.push(0xEC);
+            v
+        };
+        // Every AU carries BOTH HDR10 SEI messages, as a real HDR10 stream does.
+        let au = || {
+            let mut data = pps.clone();
+            data.extend_from_slice(&sei_nal(&[
+                sei_message(
+                    SEI_MASTERING_DISPLAY_COLOUR_VOLUME,
+                    &mastering_payload([1, 2, 3], [4, 5, 6], 7, 8, 9, 10),
+                ),
+                sei_message(SEI_CONTENT_LIGHT_LEVEL_INFO, &cll_payload(1000, 400)),
+            ]));
+            data.extend_from_slice(&idr);
+            data
+        };
+
+        let mut parser = HevcParser::new();
+        // First AU: both messages captured, so this one legitimately copies.
+        parser.parse(&make_pes(au(), Some(0)));
+        assert!(
+            parser.sei_mastering.is_some() && parser.sei_content_light.is_some(),
+            "first AU must capture both HDR10 messages"
+        );
+        // Now measure the next 50 AUs, whose SEI scan is a guaranteed no-op.
+        RBSP_COPIES.with(|c| c.set(0));
+        for i in 0..50 {
+            parser.parse(&make_pes(au(), Some(3750 * (i + 1))));
+        }
+        let copies = RBSP_COPIES.with(|c| c.get());
+        assert_eq!(
+            copies, 0,
+            "SEI RBSP must not be copied once both HDR10 messages are captured; \
+             {copies} copies over 50 access units"
+        );
+        // And the captured metadata is still surfaced on those later frames.
+        let f = parser.parse(&make_pes(au(), Some(3750 * 51)));
+        assert!(
+            f[0].coding.unwrap().hdr10().is_some(),
+            "the sticky HDR10 metadata must still ride every later frame"
+        );
+    }
+
     /// Only the mastering-display SEI (no content-light SEI) → metadata is NOT
     /// surfaced. HDR10 requires BOTH; a half-populated record is never emitted.
     #[test]
@@ -1510,6 +1652,60 @@ mod tests {
         );
     }
 
+    /// MEASURED: the keyframe parameter-set re-assert must be spliced into the
+    /// front of the already-assembled access unit IN PLACE, not built as a fresh
+    /// full-size buffer.
+    ///
+    /// It used to `prefix.extend_from_slice(&frame_data)` and then replace
+    /// `frame_data` with `prefix`, which grew a few-hundred-byte `prefix` to the
+    /// FULL access-unit size — a fresh multi-MB allocation — memcpy'd the whole
+    /// frame into it, and dropped the presized buffer. One extra whole-frame
+    /// allocation plus one extra whole-frame copy per keyframe: a 2 h UHD title at
+    /// 24 fps with a 1 s GOP is ~7,200 keyframes, ~14-28 GB of avoidable memcpy per
+    /// title. `PARAM_REASSERT_HEADROOM` exists so the splice never reallocates;
+    /// this counts the reallocations that happen, which must be zero.
+    #[test]
+    fn keyframe_param_reassert_does_not_reallocate_the_frame() {
+        fn nal(t: u8, body: &[u8]) -> Vec<u8> {
+            let mut v = vec![0x00, 0x00, 0x01];
+            v.extend_from_slice(&hevc_nal_header(t));
+            v.extend_from_slice(body);
+            v
+        }
+        let sps_body = [0x01u8; 24];
+        let pps_body = [0xA1u8, 0xA2, 0xA3];
+        let mut parser = HevcParser::new();
+        // AU1 seeds the active VPS/SPS/PPS.
+        let au1 = [
+            nal(32, &[0xAA; 12]),
+            nal(33, &sps_body),
+            nal(34, &pps_body),
+            nal(19, &[0x10; 4096]),
+        ]
+        .concat();
+        parser.parse(&make_pes(au1, Some(0)));
+
+        // A run of BARE keyframes (source omits the parameter sets), each of which
+        // takes the re-assert path. Payload sized like a real coded picture so a
+        // reallocation would be the expensive one.
+        PARAM_REASSERT_REALLOCS.with(|c| c.set(0));
+        for i in 0..30i64 {
+            let au = nal(19, &vec![0x11u8; 300_000]);
+            let f = parser.parse(&make_pes(au, Some(3600 * (i + 1))));
+            // The re-assert really happened (otherwise the count is vacuously 0).
+            assert!(
+                f[0].data.len() > 300_000,
+                "keyframe {i} must carry the re-asserted parameter sets"
+            );
+        }
+        let reallocs = PARAM_REASSERT_REALLOCS.with(|c| c.get());
+        assert_eq!(
+            reallocs, 0,
+            "the parameter-set splice must fit in the reserved headroom; \
+             {reallocs} of 30 keyframes reallocated the whole frame"
+        );
+    }
+
     /// Regression (Fight Club UHD, the real bug): id 0 is body A (→ hvcC), then
     /// redefined to B, then the title switches BACK to A. A streaming decoder
     /// (hvcC at init, in-band updates only) is sitting on B; the switch back to

src/mux/codec/truehd.rs

@@ -17,6 +17,32 @@ use super::dropgate::DropTally;
 use super::{CodecParser, Frame, PesPacket, pts_to_ns};
 use crate::mux::timeline::DISCONTINUITY_BACKSTEP_NS;
 
+/// Is `w` an MLP-family major sync — a random-access / decoder re-init point?
+///
+/// The 24-bit signature is 0xF8726F; the following byte is the STREAM TYPE:
+/// 0xBA = Dolby TrueHD (the only one Blu-ray carries), 0xBB = MLP. Both are
+/// restart points, so the keyframe / re-sync decision accepts either.
+fn is_mlp_major_sync(w: u32) -> bool {
+    (w & 0xFFFF_FFFE) == 0xF872_6FBA
+}
+
+/// Is `w` specifically the TrueHD major sync (stream type 0xBA)?
+///
+/// The 32-bit word that FOLLOWS the sync is laid out per stream type: TrueHD's
+/// `format_info` carries [31..28] audio_sampling_frequency, the 5-bit 6-channel
+/// and 13-bit 8-channel presentation channel-assignment masks; MLP's (0xBB) same
+/// word carries quantization word lengths and the MLP group sample-rate fields
+/// instead. So every site that DECODES `format_info` with the TrueHD layout must
+/// require 0xBA exactly — masking the low sync bit there read a quantization code
+/// as the rate nibble and pulled channel masks out of unrelated bits, giving the
+/// track header a wrong `SamplingFrequency` and `truehd_au_duration_ns` a wrong
+/// per-AU increment (audio drifting against video for the whole track).
+/// `None` from these helpers is the safe outcome: the caller falls back to its
+/// container-derived rate/channel count.
+fn is_truehd_major_sync(w: u32) -> bool {
+    w == 0xF872_6FBA
+}
+
 /// Duration of one TrueHD access unit in nanoseconds for the 48 kHz family
 /// (48 / 96 / 192 kHz). `access_unit_size = 40 << (ratebits & 7)` and
 /// `sample_rate = 48000 << (ratebits & 7)`; the shared shift cancels in
@@ -130,7 +156,13 @@ impl TrueHdParser {
             // The rate nibble is only trustworthy once the major sync's CRC has
             // validated (above), so capture format_info here and refine the PTS
             // cadence from it ONLY on this validated path.
-            if au.len() >= 12 {
+            // ONLY for stream type 0xBA. An MLP (0xBB) major sync's following word
+            // is not the TrueHD `format_info` layout, so decoding it as one gave a
+            // wrong rate and channel count; leaving it `None` keeps the
+            // container-derived rate, which is the honest answer.
+            if au.len() >= 12
+                && is_truehd_major_sync(u32::from_be_bytes([au[4], au[5], au[6], au[7]]))
+            {
                 format_info = Some(u32::from_be_bytes([au[8], au[9], au[10], au[11]]));
             }
         }
@@ -445,10 +477,17 @@ impl CodecParser for TrueHdParser {
                 break; // incomplete access unit, wait for more data
             }
 
+            // Restart-point question — either stream type (0xBA TrueHD, 0xBB MLP)
+            // is a decoder re-init point, so both count as a major sync here.
+            // DECODING format_info with the TrueHD layout is a separate question,
+            // gated on 0xBA alone in `au_check`.
             let is_major_sync = unit_bytes >= 8
-                && (u32::from_be_bytes([self.buf[4], self.buf[5], self.buf[6], self.buf[7]])
+                && is_mlp_major_sync(u32::from_be_bytes([
-                    & 0xFFFF_FFFE)
+                    self.buf[4],
-                    == 0xF872_6FBA;
+                    self.buf[5],
+                    self.buf[6],
+                    self.buf[7],
+                ]));
 
             // Decodability gate. MLP/TrueHD decode state persists across access
             // units, so a corrupt AU is dropped FORWARD to the next VALIDATED
@@ -597,7 +636,9 @@ pub fn truehd_channels_from_stream(data: &[u8]) -> Option<u8> {
     let mut p = 0;
     while p + 8 <= data.len() {
         let w = u32::from_be_bytes([data[p], data[p + 1], data[p + 2], data[p + 3]]);
-        if (w & 0xFFFF_FFFE) == 0xF872_6FBA {
+        // 0xBA only: `truehd_channels` reads the TrueHD `format_info` channel
+        // masks, which an MLP (0xBB) major sync does not carry.
+        if is_truehd_major_sync(w) {
             let fi = u32::from_be_bytes([data[p + 4], data[p + 5], data[p + 6], data[p + 7]]);
             return truehd_channels(fi);
         }
@@ -663,7 +704,9 @@ pub fn truehd_sync_info_from_stream(data: &[u8]) -> Option<TrueHdSyncInfo> {
     let mut p = 0;
     while p + 8 <= data.len() {
         let w = u32::from_be_bytes([data[p], data[p + 1], data[p + 2], data[p + 3]]);
-        if (w & 0xFFFF_FFFE) == 0xF872_6FBA {
+        // 0xBA only: `format_info` (and the num_substreams/Atmos nibble) are the
+        // TrueHD layout, not MLP's.
+        if is_truehd_major_sync(w) {
             let format_info =
                 u32::from_be_bytes([data[p + 4], data[p + 5], data[p + 6], data[p + 7]]);
             // num_substreams is the top nibble of the 17th sync byte (p + 16).
@@ -1447,16 +1490,53 @@ mod tests {
     // --- truehd_channels_from_stream: major-sync variant bit + scan ---
 
     #[test]
-    fn channels_from_stream_matches_variant_sync_0xfb() {
+    fn channels_from_stream_rejects_mlp_sync_0xfb() {
-        // The sync match masks the low bit: 0xF8726FBA & 0xFFFFFFFE == base, and
+        // 0xF8726FBB is the MLP stream type, NOT TrueHD (0xF8726FBA). The word
-        // 0xF8726FBB (the +1 variant) matches the same masked pattern. A stream
+        // after an MLP major sync holds quantization word lengths and the MLP
-        // carrying 0xF8726FBB must still be recognised.
+        // group sample-rate fields, not TrueHD's rate nibble plus the 6ch/8ch
+        // presentation channel-assignment masks. Decoding it with the TrueHD
+        // layout (the scan used to mask the low sync bit) reported channels out of
+        // unrelated bits; the honest answer is None so the caller keeps its
+        // container-derived count. Byte pattern below decodes as 8 channels ONLY
+        // under the TrueHD layout, so this test fails if the mask comes back.
         let mut data = vec![0x00];
         data.extend_from_slice(&0xF872_6FBBu32.to_be_bytes());
         data.extend_from_slice(&0x0000_001Fu32.to_be_bytes());
+        assert_eq!(
+            truehd_channels_from_stream(&data),
+            None,
+            "an MLP (0xBB) major sync must not be decoded as TrueHD format_info"
+        );
+        // The same bytes under the TrueHD stream type DO decode.
+        let mut data = vec![0x00];
+        data.extend_from_slice(&0xF872_6FBAu32.to_be_bytes());
+        data.extend_from_slice(&0x0000_001Fu32.to_be_bytes());
         assert_eq!(truehd_channels_from_stream(&data), Some(8));
     }
 
+    #[test]
+    fn mlp_sync_0xfb_yields_no_sample_rate_or_atmos() {
+        // Same split for the shared scan: an MLP major sync must not produce a
+        // TrueHD rate (bits 31..28 of an MLP header are a quantization code, not
+        // the rate) nor an Atmos verdict.
+        let mut data = vec![0x00];
+        data.extend_from_slice(&0xF872_6FBBu32.to_be_bytes());
+        // ratebits nibble 0x1 would decode as 96 kHz under the TrueHD layout.
+        data.extend_from_slice(&0x1000_001Fu32.to_be_bytes());
+        data.extend_from_slice(&[0x00; 12]);
+        assert!(
+            truehd_sync_info_from_stream(&data).is_none(),
+            "no TrueHD sync info from an MLP major sync"
+        );
+        assert_eq!(truehd_sample_rate_from_stream(&data), None);
+        // TrueHD stream type, identical trailing bytes → the rate IS decoded.
+        let mut data = vec![0x00];
+        data.extend_from_slice(&0xF872_6FBAu32.to_be_bytes());
+        data.extend_from_slice(&0x1000_001Fu32.to_be_bytes());
+        data.extend_from_slice(&[0x00; 12]);
+        assert_eq!(truehd_sample_rate_from_stream(&data), Some(96000));
+    }
+
     #[test]
     fn channels_from_stream_none_without_major_sync() {
         // No major sync anywhere → None, no panic, scan terminates.
@@ -1527,8 +1607,10 @@ mod tests {
 
     #[test]
     fn major_sync_variant_bit_also_keyframe() {
-        // The keyframe check masks the low bit (0xFFFF_FFFE), so the 0xF8726FBB
+        // The RESTART-POINT check masks the low sync bit, so 0xF8726FBB (MLP)
-        // variant must also be detected as a major sync.
+        // counts as a major sync too — both stream types re-init the decoder, so
+        // both are keyframes. (Only the format_info DECODE is 0xBA-only; see
+        // `channels_from_stream_rejects_mlp_sync_0xfb`.)
         let mut parser = TrueHdParser::new();
         let mut unit = make_truehd_unit(200);
         unit[4..8].copy_from_slice(&0xF872_6FBBu32.to_be_bytes());

src/mux/demux_sink.rs

@@ -220,6 +220,13 @@ struct AnnexBWriter {
     /// avcC/hvcC may declare 1 or 2, and reading those as u32-BE parses no NALs
     /// at all, so the raw prefixed bytes would be emitted as if already Annex B.
     length_size: usize,
+    /// Reused length-prefixed -> Annex-B conversion buffer. `write_frame` used to
+    /// allocate and free a whole-frame Vec per video frame; extracting the video ES
+    /// of a UHD title is ~200,000 frames of 150-400 KB, every one over the
+    /// allocator's mmap threshold, so that was ~200,000 mmap/munmap pairs plus
+    /// millions of first-touch page faults of pure overhead. Kept on the writer and
+    /// cleared per frame instead, matching what tsmux.rs already does.
+    scratch: Vec<u8>,
 }
 
 impl AnnexBWriter {
@@ -231,6 +238,7 @@ impl AnnexBWriter {
             params,
             wrote_params: false,
             length_size: nal_length_size(codec, codec_private),
+            scratch: Vec::new(),
         }
     }
 }
@@ -251,10 +259,16 @@ impl EsWriter for AnnexBWriter {
         // source of truth across all muxers — see `crate::mux::hevc`). It skips
         // zero-length NALs and drops a truncated trailing NAL without panicking,
         // rather than `break`ing on the first zero-length NAL.
-        let mut scratch = Vec::with_capacity(f.data.len() + (f.data.len() / 32) + 4);
+        // Reuse the writer's buffer rather than allocating per frame; clear()
+        // keeps the capacity, so steady state costs no allocation at all. The
+        // prefix width still comes from the record, never a hardcoded 4.
+        self.scratch.clear();
+        self.scratch.reserve(f.data.len() + (f.data.len() / 32) + 4);
+        let mut scratch = std::mem::take(&mut self.scratch);
         append_length_prefixed_as_annex_b_sized(&mut scratch, &f.data, self.length_size);
         w.write_all(&scratch)?;
         n += scratch.len();
+        self.scratch = scratch;
         Ok(n)
     }
 }

src/mux/driver.rs

@@ -825,10 +825,43 @@ fn drive_mux(
                 }
                 Err(e) => {
                     // Drain + join the consumer so its output file handle is
-                    // released, then propagate the read error.
+                    // released, then report the ROOT cause.
-                    let _ = pipe.finish_with_halt(Some(halt));
+                    //
+                    // The consumer's result used to be discarded with `let _`. A
+                    // write-side `WriteSink::apply` failure that had ALREADY killed
+                    // the output — the destination volume filling, say — was thrown
+                    // away, and only the read error surfaced: the caller diagnosed a
+                    // damaged disc and retried the rip onto the same full volume
+                    // instead of being told it was out of space. A hard write
+                    // failure precedes and explains the read error here (the
+                    // producer only reaches the next `read()` because its previous
+                    // `send` did not block on a dead consumer), so prefer it.
+                    // Halt/join-timeout are NOT root causes — those are the clean
+                    // operator-stop and wedge paths the finish stage below
+                    // translates to `completed = false` — so the read error still
+                    // wins over them.
+                    match pipe.finish_with_halt(Some(halt)) {
+                        Err(w @ (Error::Halted | Error::PipelineJoinTimeout)) => {
+                            tracing::debug!(
+                                target: "mux",
+                                write_side = %w,
+                                read_side = %e,
+                                "read failed; consumer stopped for a non-root-cause reason — reporting the read error"
+                            );
                             return Err(e);
                         }
+                        Err(w) => {
+                            tracing::error!(
+                                target: "mux",
+                                write_side = %w,
+                                read_side = %e,
+                                "read failed, but the write side had already failed — reporting the write failure as the root cause"
+                            );
+                            return Err(w.into());
+                        }
+                        Ok(_) => return Err(e),
+                    }
+                }
             }
         }
     }

src/mux/m2ts_mux/mod.rs

@@ -1151,6 +1151,15 @@ mod tests {
         })
     }
 
+    /// Every packet on `pid` (optionally requiring PUSI), in stream order.
+    fn find_all_pkts(buf: &[u8], pid: u16, pusi: bool) -> Vec<&[u8]> {
+        buf.chunks(188)
+            .filter(|p| {
+                u16::from_be_bytes([p[1] & 0x1F, p[2]]) == pid && (!pusi || (p[1] & 0x40) != 0)
+            })
+            .collect()
+    }
+
     /// Extract a PSI section (after the pointer_field) from a PUSI PSI
     /// packet: payload starts at byte 4 (no AF on PSI here), first payload
     /// byte is pointer_field, section follows.
@@ -1355,31 +1364,92 @@ mod tests {
 
     #[test]
     fn extreme_pts_does_not_overflow_and_clamps_to_33bit() {
-        // base_relative_pts widens to u128 then masks to 33 bits. An
+        // `base_relative_pts` widens to u128 then wraps the delta into 33 bits. An
-        // adversarial i64::MAX ns must not overflow and the encoded PTS must
+        // adversarial i64::MAX ns must not overflow, and the encoded PES PTS must be
-        // stay within the 33-bit field. With a single video frame the base
+        // the EXACT wrapped value.
-        // is itself, so relative PTS is 0 — proving no panic on the path.
+        //
-        let mut sink: Vec<u8> = Vec::new();
+        // The old version of this test wrote a single video frame, which by its own
-        {
+        // admission rebases to relative PTS 0 — so its only assertion was
-            let mut mux = M2tsMux::new(&mut sink);
+        // `0 < 2^33`, which cannot fail: widening the 33-bit mask (or deleting it)
+        // left the one test named for 33-bit clamping green. Two frames are needed
+        // so the second has a non-zero delta to pin, and the expected value is
+        // computed here from the spec formula rather than read back from the code
+        // under test.
+        //
+        // NOTE on the test's name: bit 32 of the PES PTS field is UNREACHABLE from
+        // this path. `base_relative_pts` interprets the delta as a signed 33-bit
+        // value and floors the entire upper half [2^32, 2^33) to 0, so anything the
+        // encoder ever receives is < 2^32. "Clamping to 33 bits" is therefore a
+        // structural property of the delta rule, not something a test can exercise;
+        // what IS pinned below is the exact encoding of the largest reachable value
+        // and the documented i64::MAX outcome.
+        // Decode the 33-bit PTS out of a PUSI video packet's PES header.
+        let decode_pts = |pkt: &[u8]| -> u64 {
+            // Payload after the AF: AF area = 1 (length byte) + af_len.
+            let af_len = pkt[4] as usize;
+            let pes = &pkt[4 + 1 + af_len..];
+            // PES: 00 00 01 E0 00 00 80 80 05 PTS[5]. PTS at pes[9..14].
+            ((((pes[9] >> 1) & 0x07) as u64) << 30)
+                | ((pes[10] as u64) << 22)
+                | (((pes[11] >> 1) as u64) << 15)
+                | ((pes[12] as u64) << 7)
+                | ((pes[13] >> 1) as u64)
+        };
         let mut frame = Vec::new();
         frame.extend_from_slice(&4u32.to_be_bytes());
         frame.extend_from_slice(&[0x40, 0x01, 0x0C, 0x01]);
+
+        // (a) A LARGE but forward delta: exact-value assertion across all 33 bits.
+        //
+        // Frame 1 at PTS 0 seeds the base to 0, so frame 2's relative PTS is its own
+        // tick value. 477_218_477 x 100_000 ns divides exactly by the 100_000/9
+        // conversion, giving 4_294_966_293 ticks — just under 2^32, i.e. the largest
+        // magnitude the signed-33-bit delta rule treats as forward progression, and
+        // a value that needs the full 3+15+15-bit PES PTS field to survive.
+        let mut sink: Vec<u8> = Vec::new();
+        {
+            let mut mux = M2tsMux::new(&mut sink);
+            mux.write_video(0, true, &frame).unwrap();
+            mux.write_video(477_218_477 * 100_000, true, &frame)
+                .unwrap();
+            mux.finish().unwrap();
+        }
+        assert_ts_well_formed(&sink);
+        let pkts = find_all_pkts(&sink, PID_VIDEO, true);
+        assert_eq!(pkts.len(), 2, "one PUSI packet per video frame");
+        let pts = decode_pts(pkts[1]);
+        assert!(pts < (1u64 << 33), "PTS stays within the 33-bit field");
+        assert_eq!(
+            pts, 4_294_966_293,
+            "the encoded PES PTS must be the exact tick value, not a truncated one"
+        );
+
+        // (b) The adversarial i64::MAX: no overflow, no panic, and the documented
+        // signed-33-bit outcome. Its tick count masks to 6_564_084_417, which is in
+        // the UPPER half of the 33-bit range, so the delta rule reads it as a frame
+        // BEFORE the base and floors it to 0 (the same rule that floors leading
+        // audio). Asserted explicitly so this is a pinned decision, not the vacuous
+        // `0 < 2^33` the old single-frame version checked.
+        let mut sink: Vec<u8> = Vec::new();
+        {
+            let mut mux = M2tsMux::new(&mut sink);
+            mux.write_video(0, true, &frame).unwrap();
             mux.write_video(i64::MAX, true, &frame).unwrap();
             mux.finish().unwrap();
         }
         assert_ts_well_formed(&sink);
-        let pkt = find_pkt(&sink, PID_VIDEO, true).unwrap();
+        let masked = (((i64::MAX as u128) * 9 / 100_000) as u64) & 0x1_FFFF_FFFF;
-        // Reach the PES PTS: payload after AF. AF area = 1 (length) + af_len.
+        assert!(
-        let af_len = pkt[4] as usize;
+            masked >= 1 << 32,
-        let pes = &pkt[4 + 1 + af_len..];
+            "i64::MAX masks into the upper (negative) half"
-        // PES: 00 00 01 E0 00 00 80 80 05 PTS[5]. PTS at pes[9..14].
+        );
-        let pts = ((((pes[9] >> 1) & 0x07) as u64) << 30)
+        let pkts = find_all_pkts(&sink, PID_VIDEO, true);
-            | ((pes[10] as u64) << 22)
+        let pts = decode_pts(pkts[1]);
-            | (((pes[11] >> 1) as u64) << 15)
-            | ((pes[12] as u64) << 7)
-            | ((pes[13] >> 1) as u64);
         assert!(pts < (1u64 << 33), "PTS stays within the 33-bit field");
+        assert_eq!(
+            pts, 0,
+            "an i64::MAX tick lands in the signed-33-bit upper half and floors to 0"
+        );
     }
 
     #[test]

src/mux/mkv.rs

@@ -681,10 +681,18 @@ pub struct MkvMuxer<W: Write + Seek> {
     cues: Vec<CuePoint>,
     frame_count: u64,
     /// Frames handed to `write_frame` that were dropped because no cluster was
-    /// open yet (a cluster only opens on a track-0 video keyframe). If this is
+    /// open yet (a cluster only opens on a track-0 video keyframe). See
-    /// non-zero at `finish()` and not a single frame was ever written, the
+    /// `write_frame` for the track-0 invariant.
-    /// caller produced an empty MKV — surfaced as an error rather than a
+    ///
-    /// silently empty file. See `write_frame` for the track-0 invariant.
+    /// The ALL-dropped case is surfaced as an error by `finish()`, but via
+    /// `frame_count == 0`, not via this counter. A PARTIAL drop — leading audio /
+    /// subtitle frames ahead of the first video IDR, or an M2TS whose PMT lists
+    /// audio before video — is normal enough not to fail the mux, but it used to
+    /// leave NO record anywhere: the field was incremented at two sites and read
+    /// nowhere (no log, no error, no accessor), so those frames vanished from the
+    /// output with `completed = true`, an empty `undelivered_streams`, and nothing
+    /// in the log. `finish()` now logs the count — that log is the field's only
+    /// reader, so do not delete it and turn this back into dead bookkeeping.
     dropped_pre_cluster: u64,
     seek_fixups: Vec<SeekPositionFixup>,
     /// Absolute file offset of the CUES SeekHead entry (a fixed 21-byte Seek
@@ -1657,6 +1665,18 @@ impl<W: Write + Seek> MkvMuxer<W> {
         if self.frame_count == 0 {
             return Err(crate::error::Error::MkvInvalid.into());
         }
+        // Partial pre-cluster drops do not fail the mux (leading audio ahead of the
+        // first video IDR is normal), but they MUST leave a record: this counter
+        // was write-only, so frames silently vanished from the output while the run
+        // reported success. See the field's doc.
+        if self.dropped_pre_cluster > 0 {
+            tracing::warn!(
+                target: "mux",
+                dropped = self.dropped_pre_cluster,
+                frames_written = self.frame_count,
+                "frames were discarded before the first cluster opened (no track-0 video keyframe had arrived yet); they are absent from the output"
+            );
+        }
         // The source declared no duration up-front (DURATION was reserved as a
         // placeholder). Derive the real runtime from the muxed timeline so the
         // Segment declares it — and so the BPS tags below can be computed.
@@ -2615,6 +2635,47 @@ mod tests {
         );
     }
 
+    /// Frames dropped before the first cluster opens must be COUNTED, and the
+    /// count must survive to `finish()` so it can be reported. The counter was
+    /// incremented at two sites and read nowhere — no log, no error, no accessor —
+    /// so a partial drop (leading audio/subtitle frames ahead of the first video
+    /// IDR, or an M2TS whose PMT lists audio before video) silently omitted those
+    /// frames from the output while the run reported `completed = true` with an
+    /// empty `undelivered_streams` and nothing in the log. `finish()` now logs it;
+    /// this pins the accounting the log depends on.
+    #[test]
+    fn frames_dropped_before_first_cluster_are_counted() {
+        let buf = Cursor::new(Vec::new());
+        let tracks = [make_video_track()];
+        let mut muxer = MkvMuxer::new(buf, &tracks, None, 60.0, &[]).unwrap();
+        // Two non-keyframes arrive before any track-0 keyframe: no cluster can be
+        // open, so both are dropped.
+        muxer
+            .write_frame(0, 0, false, &[0x01; 8], None, None)
+            .unwrap();
+        muxer
+            .write_frame(0, 1_000_000, false, &[0x02; 8], None, None)
+            .unwrap();
+        assert_eq!(
+            muxer.dropped_pre_cluster, 2,
+            "both pre-cluster frames must be counted, not silently lost"
+        );
+        assert_eq!(muxer.frame_count, 0, "neither frame was written");
+        // A keyframe then opens the cluster and is written; the drop count stands
+        // so `finish()` can still report it.
+        muxer
+            .write_frame(0, 2_000_000, true, &[0x03; 8], None, None)
+            .unwrap();
+        assert_eq!(muxer.frame_count, 1);
+        assert_eq!(
+            muxer.dropped_pre_cluster, 2,
+            "the drop count must survive to finish(), which reports it"
+        );
+        muxer
+            .finish()
+            .expect("a mux with one written frame succeeds");
+    }
+
     #[test]
     fn mkv_finish_writes_cues_element() {
         // finish() consumes self and flushes the writer, so use the

src/mux/ts.rs

@@ -53,6 +53,11 @@ pub struct PesPacket {
 /// Per-PID PES reassembly state.
 struct PesAssembler {
     pid: u16,
+    /// This PID's PES-reassembly ceiling: its share of [`MAX_PES_BUFFER_TOTAL`],
+    /// clamped to [`MAX_PES_BUFFER`]. Resolved once by `TsDemuxer::new` so the
+    /// per-PID caps sum to a bounded total no matter how many streams the disc
+    /// declares.
+    cap: usize,
     buffer: Vec<u8>,
     pts: Option<i64>,
     dts: Option<i64>,
@@ -107,10 +112,34 @@ const PES_BUFFER_INIT_CAP: usize = 16 * 1024;
 /// the next PUSI.
 const MAX_PES_BUFFER: usize = 64 * 1024 * 1024; // 64 MiB
 
+/// AGGREGATE ceiling across every tracked PID.
+///
+/// [`MAX_PES_BUFFER`] bounds each PID's buffer independently and never sees the
+/// total, while the tracked-PID count comes straight off the disc: `TsDemuxer::new`
+/// makes one [`PesAssembler`] per SELECTED stream, and the selection derives from
+/// the MPLS STN, whose per-category counts are `u8` (up to 255 each across 8
+/// categories) bounded only by the MPLS file's own bytes. A crafted MPLS declaring
+/// 100 streams on 100 distinct PIDs, plus a clip feeding each PID continuation
+/// packets (no PUSI) until just under the per-PID cap, held 100 x 64 MiB = 6.4 GiB
+/// of PES buffers at once; 1000 distinct PIDs — well inside the 8192-entry
+/// `pid_index` table — is 64 GiB.
+///
+/// So the per-PID cap is derived from this total instead: `pes_cap` (below) is
+/// `MAX_PES_BUFFER_TOTAL / tracked_pids`, clamped to `MAX_PES_BUFFER`. A real title
+/// selects a handful of streams and keeps the full 64 MiB each; only a stream count
+/// far past anything an authored disc carries is squeezed, and even then a complete
+/// HEVC/UHD access unit (1-3 MiB) still fits at ~170 PIDs. Overflow is graceful in
+/// any case — the partial PES is dropped and the assembler resyncs on the next
+/// PUSI, flagging a discontinuity.
+const MAX_PES_BUFFER_TOTAL: usize = 512 * 1024 * 1024; // 512 MiB
+
 impl PesAssembler {
-    fn new(pid: u16) -> Self {
+    /// `cap` is this PID's SHARE of [`MAX_PES_BUFFER_TOTAL`], resolved by
+    /// `TsDemuxer::new` from the tracked-PID count.
+    fn new(pid: u16, cap: usize) -> Self {
         Self {
             pid,
+            cap,
             buffer: Vec::with_capacity(PES_BUFFER_INIT_CAP),
             pts: None,
             dts: None,
@@ -155,13 +184,14 @@ impl PesAssembler {
 
     /// Append payload data to the current PES packet.
     ///
-    /// If the buffer would exceed [`MAX_PES_BUFFER`] the partial PES is
+    /// If the buffer would exceed this PID's `cap` — its share of
+    /// [`MAX_PES_BUFFER_TOTAL`], at most [`MAX_PES_BUFFER`] — the partial PES is
     /// silently dropped and the assembler is reset. Normal traffic resumes
     /// on the next PUSI; a crafted/corrupt stream that never sends one can
-    /// no longer drive unbounded allocation.
+    /// no longer drive unbounded allocation, on this PID OR in aggregate.
     fn push(&mut self, data: &[u8]) {
         if self.active {
-            if self.buffer.len().saturating_add(data.len()) > MAX_PES_BUFFER {
+            if self.buffer.len().saturating_add(data.len()) > self.cap {
                 tracing::trace!(
                     target: "mux",
                     pid = self.pid,
@@ -237,9 +267,13 @@ impl TsDemuxer {
         let table_size = (max_pid + 1).max(8192);
         let mut pid_index = vec![-1i32; table_size];
         let mut assemblers = Vec::with_capacity(pids.len());
+        // Per-PID cap = this PID's share of the AGGREGATE ceiling. Without this the
+        // caps were per-PID only and never saw the total, so a disc-declared stream
+        // list could multiply 64 MiB by its own length.
+        let pes_cap = (MAX_PES_BUFFER_TOTAL / pids.len().max(1)).min(MAX_PES_BUFFER);
         for (i, &pid) in pids.iter().enumerate() {
             pid_index[pid as usize] = i as i32;
-            assemblers.push(PesAssembler::new(pid));
+            assemblers.push(PesAssembler::new(pid, pes_cap));
         }
         Self {
             assemblers,
@@ -2229,6 +2263,58 @@ mod tests {
 
     // ── PES reassembly buffer cap (DoS hardening) ─────────────────────────
 
+    /// The per-PID PES cap must be a SHARE of an aggregate ceiling, not a flat
+    /// 64 MiB per PID that never sees the total. The tracked-PID count comes off
+    /// the disc (one assembler per selected stream, selection driven by the MPLS
+    /// STN whose per-category counts are u8), so a crafted MPLS declaring many
+    /// streams on distinct PIDs held `count x 64 MiB` of PES buffers at once —
+    /// 6.4 GiB at 100 PIDs, 64 GiB at 1000 (still inside the 8192-entry pid_index
+    /// table). With 64 tracked PIDs each share is 512 MiB / 64 = 8 MiB, so a PID
+    /// flooded with continuation packets must drop its partial PES at ~8 MiB, not
+    /// at 64 MiB.
+    #[test]
+    fn per_pid_pes_cap_is_a_share_of_an_aggregate_ceiling() {
+        let pids: Vec<u16> = (0x1000..0x1040).collect(); // 64 PIDs
+        assert_eq!(pids.len(), 64);
+        let mut demux = TsDemuxer::new(&pids);
+
+        let expected_share = MAX_PES_BUFFER_TOTAL / 64;
+        assert!(
+            expected_share < MAX_PES_BUFFER,
+            "the test is only meaningful when the share is below the per-PID cap"
+        );
+        // The caps must sum to the aggregate ceiling, never to 64 x 64 MiB.
+        let total: usize = demux.assemblers.iter().map(|a| a.cap).sum();
+        assert!(
+            total <= MAX_PES_BUFFER_TOTAL,
+            "per-PID caps must sum within the aggregate ceiling: {total} > {MAX_PES_BUFFER_TOTAL}"
+        );
+
+        // Behavioural: flood ONE PID with continuation packets and confirm the
+        // partial PES is dropped at its share, not at MAX_PES_BUFFER.
+        let pid = pids[0];
+        let mut pes_start = vec![0x00, 0x00, 0x01, 0xE0, 0x00, 0x00, 0x80, 0x00, 0x00];
+        pes_start.extend_from_slice(&[0xAB; 10]);
+        demux.feed(&es_packet_exact(pid, true, &pes_start));
+        let payload = [0xCCu8; 184];
+        let cont_pkt = data_packet(pid, false, &payload);
+        let mut high_water = 0usize;
+        for _ in 0..(expected_share / 184 + 64) {
+            demux.feed(&cont_pkt);
+            let idx = demux.pid_index[pid as usize] as usize;
+            high_water = high_water.max(demux.assemblers[idx].buffer.len());
+        }
+        assert!(
+            high_water <= expected_share,
+            "a flooded PID must be capped at its share ({expected_share}), \
+             not at the flat per-PID cap; high water was {high_water}"
+        );
+        assert!(
+            high_water > expected_share / 2,
+            "sanity: the flood must actually have filled the share, got {high_water}"
+        );
+    }
+
     #[test]
     fn pes_buffer_cap_resets_on_overflow_and_recovers_on_next_pusi() {
         // Feed continuation-only packets that would exceed MAX_PES_BUFFER if

src/mux/tsmux.rs

@@ -265,14 +265,22 @@ impl<W: Write> TsMuxer<W> {
 
         // Video PES may be unbounded (length 0); a 0xBD private_stream_1
         // PES must carry a bounded length, so split oversized audio/sub
-        // access units into multiple PES packets. Each emitted PES carries
+        // access units into multiple PES packets.
-        // the same PTS and starts on its own PUSI packet (only the keyframe
+        //
-        // RAI rides the first packet of the first PES).
+        // ONLY THE FIRST emitted PES carries the PTS. ISO/IEC 13818-1 §2.4.3.7
+        // puts the PTS in the header of the PES packet containing the FIRST byte of
+        // the access unit; every chunk used to repeat it, so on read-back a demuxer
+        // (which treats each PUSI as a new access unit) received the second half of
+        // e.g. an oversized full-screen PGS display set as an independent segment at
+        // the SAME timestamp, and the display set was emitted as two blocks with
+        // identical timestamps instead of one. Each PES still necessarily starts on
+        // its own PUSI packet — that is what delimits a PES — but only the keyframe
+        // RAI rides the first packet of the first PES.
         //
         // The write result is held rather than `?`-propagated so the conversion
         // buffer goes back into `self` on every path.
         let res = if is_video || es_data.len() <= MAX_BD_PES_PAYLOAD {
-            self.write_pes_chain(track, pid, pts_90k, is_video, keyframe, es_data)
+            self.write_pes_chain(track, pid, Some(pts_90k), is_video, keyframe, es_data)
         } else {
             let mut first_pes = true;
             let mut res = Ok(());
@@ -280,7 +288,7 @@ impl<W: Write> TsMuxer<W> {
                 res = self.write_pes_chain(
                     track,
                     pid,
-                    pts_90k,
+                    first_pes.then_some(pts_90k),
                     is_video,
                     keyframe && first_pes,
                     chunk,
@@ -309,7 +317,7 @@ impl<W: Write> TsMuxer<W> {
         &mut self,
         track: usize,
         pid: u16,
-        pts_90k: u64,
+        pts_90k: Option<u64>,
         is_video: bool,
         keyframe: bool,
         es_data: &[u8],
@@ -434,7 +442,14 @@ impl<W: Write> TsMuxer<W> {
 }
 
 /// Build a PES packet header for a BD stream.
-fn build_pes_header(pid: u16, pts_90k: u64, data_len: usize) -> Vec<u8> {
+/// `pts_90k` is `None` for a CONTINUATION PES packet — one carrying the rest of an
+/// access unit that was too large for a single bounded-length private_stream_1 PES.
+/// ISO/IEC 13818-1 §2.4.3.7 puts the PTS in the header of the PES packet that
+/// contains the FIRST byte of the access unit; repeating it on the continuations
+/// makes each of them look like a new access unit at the same timestamp, so a
+/// demuxer re-reading the stream splits one display set into two blocks with
+/// identical timestamps.
+fn build_pes_header(pid: u16, pts_90k: Option<u64>, data_len: usize) -> Vec<u8> {
     use crate::consts::pes_stream_id;
     // Determine stream_id from PID range
     let stream_id: u8 = if is_video_pid(pid) {
@@ -443,7 +458,8 @@ fn build_pes_header(pid: u16, pts_90k: u64, data_len: usize) -> Vec<u8> {
         pes_stream_id::PRIVATE_STREAM_1 // audio, PGS subtitle, or default
     };
 
-    let pes_data_len = data_len + 8; // 3 header bytes + 5 PTS bytes + data
+    // 3 optional-header bytes + 5 PTS bytes (when present) + data.
+    let pes_data_len = data_len + if pts_90k.is_some() { 8 } else { 3 };
     let mut header = Vec::with_capacity(14);
 
     // Start code: 00 00 01 stream_id
@@ -465,8 +481,14 @@ fn build_pes_header(pid: u16, pts_90k: u64, data_len: usize) -> Vec<u8> {
         header.push(len as u8);
     }
 
-    // Flags: 10xx xxxx — MPEG-2, PTS present
+    // Flags: 10xx xxxx — MPEG-2
     header.push(0x80); // marker bits
+    let Some(pts_90k) = pts_90k else {
+        // Continuation packet: PTS_DTS_flags = 00, no optional fields.
+        header.push(0x00);
+        header.push(0);
+        return header;
+    };
     header.push(0x80); // PTS present
 
     // PES header data length
@@ -863,10 +885,15 @@ mod tests {
         let mut out = Vec::new();
         for p in packets.iter().filter(|p| p.pid == pid) {
             if p.pusi {
-                // Skip the 14-byte PES header (3 startcode + 1 stream_id +
+                // Read the PES header's own length rather than assuming one:
-                // 2 length + 2 flags + 1 hdr_len + 5 PTS).
+                // 6 bytes (startcode + stream_id + length) + 3 optional-header
-                assert!(p.payload.len() >= 14, "PUSI payload holds a PES header");
+                // bytes + PES_header_data_length. A CONTINUATION PES carries no PTS
-                out.extend_from_slice(&p.payload[14..]);
+                // (ISO/IEC 13818-1 §2.4.3.7), so its header is 9 bytes, not 14 —
+                // this helper used to hardcode 14 and so silently depended on every
+                // split chunk repeating the PTS.
+                assert!(p.payload.len() >= 9, "PUSI payload holds a PES header");
+                let hdr = 9 + p.payload[8] as usize;
+                out.extend_from_slice(&p.payload[hdr..]);
             } else {
                 out.extend_from_slice(&p.payload);
             }
@@ -874,6 +901,15 @@ mod tests {
         out
     }
 
+    /// PTS_DTS_flags of every PUSI PES header on `pid`, in order.
+    fn pes_pts_flags(packets: &[TsPacket], pid: u16) -> Vec<u8> {
+        packets
+            .iter()
+            .filter(|p| p.pid == pid && p.pusi)
+            .map(|p| (p.payload[7] >> 6) & 0x03)
+            .collect()
+    }
+
     /// A non-NAL codec must pass the ES through byte-for-byte: MPEG-2 and
     /// VC-1 are not NAL-based, so their ES already IS the wire format and
     /// `length_prefixed_to_annex_b` would mangle it.
@@ -1179,6 +1215,44 @@ mod tests {
         assert_eq!(got, big, "split audio reassembles byte-for-byte");
     }
 
+    /// ISO/IEC 13818-1 §2.4.3.7: the PTS belongs in the header of the PES packet
+    /// that contains the FIRST byte of the access unit. An oversized
+    /// private_stream_1 access unit is split across several PES packets, and every
+    /// one of them used to carry the SAME PTS (PTS_DTS_flags = 0b10) even though
+    /// only the first holds the start of the AU. On read-back a demuxer treats each
+    /// PUSI as a new access unit, so the second half of e.g. a full-screen PGS
+    /// display set arrived as an independent segment at an identical timestamp and
+    /// the display set was emitted as TWO blocks with the same timestamp instead of
+    /// one. Only the first PES may carry a PTS; the continuations must set
+    /// PTS_DTS_flags = 0b00.
+    #[test]
+    fn split_access_unit_carries_pts_only_on_the_first_pes() {
+        // Three PES worth of ES so there are two continuations to check.
+        let big: Vec<u8> = (0..(2 * MAX_BD_PES_PAYLOAD + 3000))
+            .map(|i| (i & 0xFF) as u8)
+            .collect();
+        let mut sink: Vec<u8> = Vec::new();
+        {
+            let mut mux = TsMuxer::new(&mut sink, &[AUDIO_PID]);
+            mux.write_frame(0, 1_000_000_000, false, &big).unwrap();
+            mux.finish().unwrap();
+        }
+        let packets = parse_bd_ts(&sink);
+        let flags = pes_pts_flags(&packets, AUDIO_PID);
+        assert_eq!(flags.len(), 3, "the AU must split into three PES packets");
+        assert_eq!(
+            flags,
+            vec![0b10, 0b00, 0b00],
+            "only the PES containing the first byte of the access unit may carry a PTS"
+        );
+        // And the split is still lossless with the shorter continuation headers.
+        assert_eq!(
+            reassemble_es(&packets, AUDIO_PID),
+            big,
+            "split audio still reassembles byte-for-byte"
+        );
+    }
+
     /// MEASURED: the Annex-B conversion buffer must be REUSED across video
     /// frames, not allocated per frame. Both the allocation's address and its
     /// capacity are unchanged after the second and third same-sized frames — if

src/mux/videomap.rs

@@ -20,7 +20,7 @@
 //! different output format would be a DIFFERENT sink reusing this same model,
 //! not a pluggable encoder here.
 
-use crate::disc::{ColorSpace, DiscTitle, FrameRate, Stream as DiscStream, VideoStream};
+use crate::disc::{ColorSpace, DiscTitle, Stream as DiscStream, VideoStream};
 use crate::mux::codec::PictureInfo;
 use crate::mux::codec::coding::{CodingType, FieldOrder};
 use crate::pes::{PesFrame, SourcePos};
@@ -268,13 +268,6 @@ fn display_aspect_ratio(v: &VideoStream, w: u32, h: u32) -> (u32, u32) {
     }
 }
 
-/// The title's nominal frame rate as a fraction — the single mapping site reused
-/// by the header builder. (Retained as the canonical accessor.)
-#[allow(dead_code)]
-fn frame_rate_fraction(fr: FrameRate) -> (u32, u32) {
-    fr.as_fraction()
-}
-
 /// One per-picture index record, distilled from a video [`PesFrame`]
 /// (`docs/FVI_FORMAT.md` §7).
 ///

src/sector/prefetched.rs

@@ -403,10 +403,32 @@ impl Drop for PrefetchShell {
 
 impl Drop for PrefetchedSectorSource {
     fn drop(&mut self) {
-        // Dropping the receiver closes the channel, which makes the
+        // Drop the channel endpoints BEFORE joining the producer.
-        // next producer `send` return Err and exits the loop. Joining
+        //
-        // here gives us a deterministic shutdown — no detached thread
+        // `rx` and `recycle_tx` are sibling fields, so they are dropped only
-        // can outlive the source.
+        // AFTER this `Drop::drop` body returns. Joining first therefore joined
+        // while both endpoints were still alive: a producer parked in the plain
+        // blocking `tx.send(Ok(buf))` (no timeout, so the `Halt` is never
+        // re-polled) never observed a disconnect and never returned, and the
+        // dropping thread blocked in `join()` forever. Any source dropped before
+        // its extents were drained — an error path, an operator stop, or a
+        // consuming crate using the public `new` + direct `read_sectors` — hit a
+        // permanent two-thread deadlock. `BytePrefetcher::drop` already had this
+        // shape; see `drop_undrained_source_joins_cleanly`.
+        //
+        // The endpoints are moved out via `mem::replace` with already-disconnected
+        // stand-ins (each stand-in's peer is dropped immediately), which drops the
+        // real ones here and needs neither `Option` fields nor `unsafe` — and
+        // leaves `into_channels`'s `ptr::read` moves untouched.
+        let (dead_tx, dead_rx) = bounded::<Batch>(0);
+        drop(dead_tx);
+        drop(std::mem::replace(&mut self.rx, dead_rx));
+        let (dead_send, dead_recv) = bounded::<Vec<u8>>(0);
+        drop(dead_recv);
+        drop(std::mem::replace(&mut self.recycle_tx, dead_send));
+        // Now the producer's next `send`/`recv` returns Err and its loop exits;
+        // joining gives a deterministic shutdown — no detached thread can outlive
+        // the source.
         if let Some(h) = self.producer.take() {
             let _ = h.join();
         }
@@ -615,6 +637,42 @@ mod tests {
         }
     }
 
+    /// Regression: dropping a `PrefetchedSectorSource` DIRECTLY — the
+    /// public `new` + documented direct-read path, and any error/halt
+    /// exit before the extents are drained — must join the producer
+    /// cleanly. `Drop::drop` used to `join()` while the struct still
+    /// held `rx` and `recycle_tx` (sibling fields drop only AFTER
+    /// `Drop::drop` returns), so a producer parked in the plain blocking
+    /// `tx.send(Ok(buf))` never saw a disconnect and the dropping thread
+    /// blocked in `join()` forever — a permanent two-thread deadlock that
+    /// cancelling the `Halt` could not escape, because that `send` has no
+    /// timeout and never re-polls the token.
+    ///
+    /// 300 sectors at batch=3 is 100 batches against a forward channel of
+    /// depth `PREFETCH_CHANNEL_DEPTH` (2), so the producer is guaranteed
+    /// to be blocked in `send` by the time the drop runs. Mirrors
+    /// `byte_prefetcher::drop_endless_prefetcher_joins_cleanly`, whose
+    /// `Drop` already had the correct shape.
+    #[test]
+    fn drop_undrained_source_joins_cleanly() {
+        with_watchdog(Duration::from_secs(10), || {
+            let extents = vec![Extent {
+                start_lba: 0,
+                sector_count: 300,
+            }];
+            let halt = Halt::new();
+            let pf = PrefetchedSectorSource::new(
+                PatternSource { capacity: 9999 },
+                extents,
+                3,
+                Some(halt.clone()),
+            )
+            .expect("spawn");
+            // Drop without draining a single batch — the old Drop deadlocked here.
+            drop(pf);
+        });
+    }
+
     /// The CRITICAL regression: after `into_channels`, dropping the
     /// returned forward receiver + recycle sender must let the producer
     /// observe disconnection and exit, so dropping the `PrefetchShell`
@@ -1002,11 +1060,11 @@ mod tests {
             let pf = PrefetchedSectorSource::new(src, extents, 3, None).expect("spawn");
             // 9 + 6 + 3 = 18, independent of inner source capacity.
             assert_eq!(pf.capacity_sectors(), 18);
-            // Release the producer without draining: peel the channels
+            // Release the producer without draining via the production
-            // and drop them so the producer observes disconnection
+            // zero-copy path: peel the channels and drop them so the
-            // (dropping `pf` directly would join while still holding the
+            // producer observes disconnection. (A direct `drop(pf)` is
-            // channels → deadlock; the production drain path always uses
+            // also safe now — see `drop_undrained_source_joins_cleanly`
-            // into_channels).
+            // — but the mux always uses into_channels.)
             let (rx, recycle_tx, shell) = pf.into_channels();
             drop(rx);
             drop(recycle_tx);
@@ -1042,9 +1100,9 @@ mod tests {
                 u32::MAX,
                 "summed total must saturate at u32::MAX, not wrap"
             );
-            // Release the producer via into_channels + drop (a direct
+            // Release the producer via into_channels + drop, the production
-            // drop of `pf` would join while still holding the channels →
+            // zero-copy path. (A direct `drop(pf)` also joins cleanly now —
-            // deadlock against the still-running EndlessZeroSource).
+            // see `drop_undrained_source_joins_cleanly`.)
             let (rx, recycle_tx, shell) = pf.into_channels();
             drop(rx);
             drop(recycle_tx);

src/udf.rs

@@ -1315,6 +1315,15 @@ impl BufferedSectorReader<'_> {
         // Cap to 8192 sectors (16 MiB) so a disc-controlled ad_len cannot
         // drive a multi-hundred-MiB allocation before any sectors are read.
         let count = count.min(8192);
+        // Clamp the range to the u32 LBA space. `start_lba` comes straight off
+        // the disc (ECMA-167 §14.1 `extent_location` is an unconstrained Uint32),
+        // so a metadata partition declared near the top of the space made
+        // `start_lba + offset` below overflow: an 'attempt to add with overflow'
+        // panic in debug inside the public `Disc::scan`, and in release a wrap to
+        // a low LBA that filled the sliding cache with a completely different
+        // region while `cache_start` still claimed the high one. Sectors past
+        // `u32::MAX` cannot be addressed at all, so dropping them loses nothing.
+        let count = count.min(u32::MAX - start_lba);
         let total = count as usize * 2048;
         self.cache.resize(total, 0);
         let mut offset = 0u32;
@@ -1359,6 +1368,16 @@ impl BufferedSectorReader<'_> {
         let mut done: u64 = 0;
         let mut hb = crate::progress::Heartbeat::new("udf_prefetch");
         for &(start, count) in ranges {
+            // Clamp each range to the u32 LBA space before walking it. Ranges come
+            // from `collect_file_ranges`, whose allocation descriptors are
+            // unconstrained ECMA-167 §14.14.1 Uint32s, and nothing bounds
+            // `start + count`; a range within one batch of the top of the space
+            // made `start + offset` and `start + offset + i` below overflow —
+            // a debug panic inside the public `Disc::scan`, and in release a wrap
+            // that seeded the PERMANENT cache with this file's bytes keyed at low
+            // LBAs, so every later single-sector read of those LBAs (the AVDP/VDS
+            // re-reads) silently parsed the wrong sector.
+            let count = count.min(u32::MAX - start);
             let mut offset = 0u32;
             while offset < count {
                 hb.tick(done, total);
@@ -1409,8 +1428,13 @@ impl SectorSource for BufferedSectorReader<'_> {
                 buf[..2048].copy_from_slice(data);
                 return Ok(2048);
             }
-            // Check sliding cache
+            // Check sliding cache. Tested as a DISTANCE from `cache_start`, not
-            if lba >= self.cache_start && lba < self.cache_start + self.cache_sectors {
+            // as `cache_start + cache_sectors`: `cache_start` is a disc-controlled
+            // LBA (and the batch-read path below sets it verbatim), so the sum
+            // overflowed for a window near `u32::MAX` — a debug panic inside
+            // `SectorSource::read_sectors`, and in release a wrap to a small value
+            // that silently disabled the cache.
+            if lba >= self.cache_start && lba - self.cache_start < self.cache_sectors {
                 let offset = (lba - self.cache_start) as usize * 2048;
                 buf[..2048].copy_from_slice(&self.cache[offset..offset + 2048]);
                 return Ok(2048);
@@ -1487,7 +1511,14 @@ mod tests {
             }
             for i in 0..count as u32 {
                 let off = i as usize * 2048;
-                let s = self.sectors.get(&(lba + i)).copied().unwrap_or([0u8; 2048]);
+                // `wrapping_add`: the near-`u32::MAX` overflow regressions below
+                // hand this harness LBAs at the top of the space on purpose, and
+                // the harness's own bookkeeping must not be what panics.
+                let s = self
+                    .sectors
+                    .get(&lba.wrapping_add(i))
+                    .copied()
+                    .unwrap_or([0u8; 2048]);
                 buf[off..off + 2048].copy_from_slice(&s);
             }
             Ok(need)
@@ -2332,6 +2363,71 @@ mod tests {
         );
     }
 
+    /// `prefetch` advances the read LBA with `start_lba + offset`. A UDF whose
+    /// metadata descriptor declares a partition near the top of the LBA space
+    /// (ECMA-167 §14.1: `extent_location` is an unconstrained Uint32) drove that
+    /// add past `u32::MAX` — an 'attempt to add with overflow' panic in debug
+    /// inside the public `Disc::scan`, and in release a wrap to a low LBA that
+    /// filled the sliding cache with a completely different region while
+    /// `cache_start` still claimed the high one.
+    #[test]
+    fn prefetch_near_u32_max_does_not_overflow() {
+        let mut inner = MapReader::new();
+        let mut br = BufferedSectorReader::new(&mut inner, 60);
+        // start + count = 0xFFFF_FFC0 + 200 > u32::MAX: the second batch
+        // iteration evaluates 0xFFFF_FFC0 + 60.
+        br.prefetch(0xFFFF_FFC0, 200);
+        // Nothing read (MapReader serves no sector here), but crucially the
+        // walk must never form an LBA above u32::MAX.
+        assert!(
+            br.cache_start.checked_add(br.cache_sectors).is_some(),
+            "cache window must stay inside the u32 LBA space: {} + {}",
+            br.cache_start,
+            br.cache_sectors
+        );
+    }
+
+    /// `prefetch_ranges` walks each disc-derived `(start_lba, sector_count)`
+    /// range with `start + offset + i`. `collect_file_ranges` permits any LBA up
+    /// to `u32::MAX` (ECMA-167 §14.14.1 allocation descriptors are unconstrained
+    /// Uint32s) and nothing bounds `start + count`, so a range within one batch
+    /// of the top of the space overflowed: debug panic inside `Disc::scan`, or in
+    /// release a wrap that seeded the PERMANENT cache with this file's bytes keyed
+    /// at LBA 0 — every later single-sector read of those low LBAs (the AVDP/VDS
+    /// re-reads) then returned the wrong sector.
+    #[test]
+    fn prefetch_ranges_near_u32_max_does_not_overflow() {
+        let mut inner = MapReader::new();
+        let mut br = BufferedSectorReader::new(&mut inner, 60);
+        br.prefetch_ranges(&[(0xFFFF_FFF0, 512)]);
+        // Every key the permanent cache holds must be a real LBA, i.e. inside
+        // the declared range — never a wrapped low sector.
+        for &lba in br.prefetched.keys() {
+            assert!(
+                lba >= 0xFFFF_FFF0,
+                "prefetch_ranges wrapped a near-u32::MAX LBA to {lba}"
+            );
+        }
+    }
+
+    /// The sliding-cache hit test computed `cache_start + cache_sectors`. Once
+    /// `cache_start` is a disc-controlled LBA near `u32::MAX` (set by the batch
+    /// read below) that add overflowed: debug panic inside
+    /// `SectorSource::read_sectors`, release wrap to a small value that silently
+    /// disabled the cache.
+    #[test]
+    fn cache_hit_test_near_u32_max_does_not_overflow() {
+        let mut inner = MapReader::new();
+        let mut br = BufferedSectorReader::new(&mut inner, 60);
+        let mut buf = [0u8; 2048];
+        // First read seeds cache_start = 0xFFFF_FFF5, cache_sectors = 60.
+        br.read_sectors(0xFFFF_FFF5, 1, &mut buf, true)
+            .expect("seed read");
+        // Second read of the same LBA evaluates the hit test: 0xFFFF_FFF5 + 60.
+        br.read_sectors(0xFFFF_FFF5, 1, &mut buf, true)
+            .expect("cache hit test must not overflow");
+    }
+
     /// Build a 2048-byte directory sector containing `count` minimal file FIDs.
     ///
     /// Each FID uses a 2-byte name (compression-id `8` + `b'A'`), so l_fi=2