# `bnb` — design rationale
`bnb` is an owned, bit-aware binary codec: arbitrary-width integers, bitfields,
enum⇄integer mappings, flag sets, a required-by-default builder, and a unified
`#[bin]` whole-message codec — one crate, integer-backed, fast. This document is the
design rationale: the problem it solves, the shape of the solution, and the decisions
behind it. For runnable walkthroughs see the [`bnb::guide`] module; for credit see
[`ACKNOWLEDGMENTS.md`](ACKNOWLEDGMENTS.md).
[`bnb::guide`]: https://docs.rs/bnb/latest/bnb/guide/
## 1. The problem
From-scratch protocol implementations need to pack and parse binary layouts that are,
all at once:
- **bit-aware** — fields narrower than a byte (a 4-bit opcode), odd widths (a 12-bit
length), and fields that straddle byte boundaries (a 108-bit payload at bit 8);
- **order-explicit** — independent control of *bit order* (does the first field land
in the high or low bits — MSB-first matches RFC diagrams) and *byte order*
(big/little), because protocols mix them;
- **fast** — shift/mask on machine integers, no `bitvec`-style bit-vector backing;
- **dual-use** — RFC-correct by default, but deliberately able to emit and accept
non-conformant data for fuzzing, red-teaming, and interop testing.
No single existing crate delivers all four. The capabilities were previously spread
across a stack of overlapping helpers — `arbitrary-int` (sub-byte ints),
`modular-bitfield`/`bitfield-struct`/`bitbybit` (bitfield packing, each with a
different bit-order convention), `num_enum` (enum⇄int) — glued to a byte-oriented
codec by hand. The glue was the pain: a byte-oriented `Read + Seek` codec has no idea
bit fields exist, so every bitfield needed `map`-style conversion glue, and the same
16-bit `opcode/flags/rcode` header would be implemented three different ways in three
protocols. `bnb` collapses that stack into one coherent, bit-native crate.
## 2. Inspiration, and why an owned codec
`bnb` is modeled on [`binrw`](https://github.com/jam1garner/binrw): its declarative,
bidirectional read/write derive and its `#[br]`/`#[bw]`/`#[brw]` attribute vocabulary
are the design `#[bin]` echoes, so a `binrw` user is immediately at home. The
arbitrary-width integers, bitfield packing, and enum mapping draw on `arbitrary-int`,
`modular-bitfield`/`bitfield-struct`/`bitbybit`, and `num_enum` respectively. `bnb`
shares no code with any of them — it is a from-scratch implementation.
The landscape that motivated building rather than composing:
| byte-oriented derive codec (`binrw`-style) | none (needs glue) | fast, byte-aligned | mature stream machinery, but bit-blind |
| bit-aware derive codec (`deku`-style) | native | slower (`bitvec` backing taxes even byte paths) | one crate, but pays the bit-vector cost |
| parser combinators (`nom`) | via `nom::bits` | fast | parse-only, not declarative bidirectional |
| transmute (`zerocopy`) | none | fastest | fixed `repr(C)` only; no variable-length/bit packing |
The two declarative bidirectional options sit on opposite horns: byte-native is fast
but bit-blind; bit-native is bit-aware but pays a bit-vector tax. The decisive case is
a layout like a **DMR burst** — 264 bits = `108 | 48 | 108`, none byte-aligned. A byte
cursor can only address byte boundaries, so such a field forces hand-rolled backward
seeks and nibble shifts; a bit-vector codec handles it but is slow everywhere. `bnb`
takes a third path: a **bit cursor over machine integers** — bit-aware *and* fast
(shift/mask, no `bitvec`). That capability is the reason the codec is owned rather than
layered on a byte-oriented one.
## 3. Architecture
### 3.1 One keystone: the `Bits` trait
Everything composes through one trait — a value that occupies a fixed number of bits:
```rust
pub trait Bits: Copy {
const BITS: u32;
fn into_bits(self) -> u128; // the value in the low BITS bits of a u128
fn from_bits(raw: u128) -> Self; // reconstruct from the low BITS bits
}
```
`bool`, the primitive unsigned integers, the `u1`..`u127` arbitrary-width integers, and
every type the macros generate (`#[bitfield]`, `#[derive(BitEnum)]`, `#[bitflags]`) all
implement `Bits`. Because the unit of composition is "a value of N bits," a 5-bit enum
nests in a 16-bit bitfield which nests in a byte-aligned `#[bin]` message — without any
glue, and with widths checked by the compiler.
A `Bits` value is the unit of bit-*packing*, but it is also a unit of *stream* coding:
every `Bits` type additionally implements the message codec traits
(`BitDecode`/`BitEncode`/`FixedBitLen`) as thin delegations to reading/writing its bits.
So `#[bin]` decodes, encodes, and sizes every field — a `Bits` leaf *or* a nested `#[bin]`
message — through one uniform interface, with no marker to disambiguate the two (see §8).
### 3.2 Two crates
A proc-macro crate cannot also export runtime items, so:
- **`bnb`** — the runtime: `int` (`UInt<T, N>` + `u1..u127`), `field` (`Bits`,
`Bitfield`, `BitOrder`, `ByteOrder`), `error`, `builder`, and `bitstream` (the codec
runtime — cursors, traits, the I/O ladder). Re-exports the macros, so users depend
only on `bnb`.
- **`bnb-macros`** — `#[bitfield]`, `#[derive(BitEnum)]`, `#[bitflags]`,
`#[derive(BitsBuilder)]`, the low-level `#[derive(BitDecode/BitEncode)]`, and `#[bin]`.
### 3.3 The load-bearing macro idea: const-eval widths
A `#[bitfield]` cannot know the numeric width of a field whose type is another
bitfield or enum — that width lives in `<T as Bits>::BITS`, resolved by the compiler.
So instead of computing offsets itself, the macro emits **const expressions**
(`<T as Bits>::BITS`, cumulative sums, offset/mask arithmetic) that the compiler
evaluates during const-eval; the generated accessors then shift/mask the single
backing integer. The same principle sizes a fixed message's `FixedBitLen::BIT_LEN`.
The proc-macro never guesses a width — the compiler does the arithmetic, and an
impossible layout is a compile error rather than a silent miscompile.
## 4. Field types and macros
- **`u1`..`u127`** (`UInt<T, N>`) — range-checked sub-byte integers backed by the
smallest sufficient primitive. Checked (`try_new`), panicking (`new`), and masking
(`from_raw`) constructors.
- **`#[bitfield]`** — packs typed `Bits` fields into one backing integer, with
`bits = msb|lsb` and `bytes = be|le` as independent knobs, and inferred /
`#[bits(N)]` / `#[bits(A..=B)]` (manual range) width forms. Generates getters,
immutable `with_*` setters, in-place `set_*`, and allocation-free `*_bytes`
conversions.
- **`#[derive(BitEnum)]`** — enum ⇄ integer at a chosen width. A `#[catch_all]`
variant preserves unknown values (the dual-use convention); without one, the enum
must cover its whole width or be marked `#[bit_enum(uN, closed)]` (otherwise it is a
compile error, since the infallible decode path would have nowhere to put an unknown
value). A byte-aligned width also gets `num_enum`-parity `From`/`TryFrom`.
- **`#[bitflags]`** — a named set of single-bit flags with full set algebra
(operators, `contains`/`iter`, per-flag accessors), dual-use retain-vs-truncate.
- **`#[derive(BitsBuilder)]`** — a required-by-default builder whose `build()` names
the first unset field, closing the gap the infallible `with_*` setters leave (a
field you forget is silently zero). `#[builder(default)]` / `#[builder(default = e)]`
opt a field out.
## 5. The `#[bin]` codec
`#[bin]` folds the read codec, the write codec, and a required-by-default builder over
one struct, generating the decode entry points (`decode`/`decode_all`/`decode_iter`/
`peek`/`decode_exact`), the encode entry points (`to_bytes`, plus `to_canonical_bytes` for a
message that has a `reserved`/`calc` field — see §5.2), the `encode(writer)` convenience
(and `BitEncode::bit_encode` for writing into a `Sink`), and construction
(`Type::new(fields…)`, `Type::builder()`). Fields are read and written at arbitrary bit
offsets, so the same attribute handles byte-aligned headers and sub-byte frames, and any
`Bits` type *or* nested `#[bin]` message drops in as a field — both decoded, encoded, and
sized through one uniform codec path, with no marker (§8).
**Struct-level options:** `big`/`little`, `bit_order = msb|lsb`, `magic = <expr>`
(a leading constant verified on read, emitted on write — any `Bits` value, so it can be
sub-byte), `read_only`/`write_only`, `no_builder`, `forward_only`, `ctx(name: Ty, …)`,
and `validate = <path>`.
**Field directives** (`#[br]`/`#[bw]`, the inherited vocabulary): `count`, `ctx { … }`,
`temp` + `calc`, `if(…)`, `map`/`try_map` (+ inverse `bw(map)`),
`parse_with`/`write_with`, `ignore`, `pad_*`/`align_*`, `restore_position`, and
`#[reserved]`/`#[reserved_with(…)]`.
`#[bin]` lowers to `#[derive(BitDecode, BitEncode, BitsBuilder)]`; those bare derives
are the codec without the builder/`#[bin]` sugar, and they carry a **right-tool guard**
— a const-eval assert that rejects an all-byte-aligned struct (the cursor never leaves
byte boundaries, so `#[bin]` is the better tool, and a sub-byte run that fills one
integer wants `#[bitfield]`). The guard is advisory steering, with
`#[bit_stream(allow_byte_aligned)]` as the escape hatch; `#[bin]` always suppresses it.
### 5.1 Tagged-union enums
`#[bin]` also applies to an *enum* — a protocol union that selects one of several
payloads. The design keeps two concerns deliberately **orthogonal**, because protocols
mix them:
- **`magic`** — a wire constant that is *read and written* (a byte string like
`b"IHDR"`, or a width-suffixed integer like `0x01u16`). Under magic dispatch it *is*
the discriminant; combined with a tag it is a post-selection signature.
- **`tag`** — a read-only **selector** drawn from `ctx`, never on the wire. The parent
passes it down (`#[br(ctx { … })]`); `tag()` recovers it to drive a no-drift `calc`.
The two compose, and may be mixed in one enum (tag priority, then magic) — the same
wire-constant-vs-selector split `#[bin]` draws on the struct side. The dual-use rule
carries over verbatim: a `#[catch_all]` variant preserves an unknown discriminant
rather than rejecting it; only an explicitly *closed* magic set errors. Variable-width
byte-string magics reuse the same [`SeekSource`](#6-the-io-ladder) capability the
positioning directives need — the bit cursor does the peeking, not a parallel mechanism.
The worked encodings live in the `bnb::guide::dispatch` page.
The encode model and construction surface below (§5.2) are **struct-only** — a
tagged-union enum encodes verbatim (no `to_canonical_bytes`/`encode_mode`/`validate`/`new`).
Those are properties of a concrete record; an enum's per-variant payloads define them, not
the union.
### 5.2 Encode model, construction, and validity
A message has two **encode forms**. `to_bytes` is **verbatim** — exactly what's stored, so
`decode → to_bytes` round-trips byte-for-byte and a deliberately-wrong field goes on the
wire as-is (dual-use). `to_canonical_bytes` is **canonical** — `reserved` fields written as
their spec value and `calc` fields recomputed, always spec-compliant. The two differ only
when a message has a `reserved` or non-`temp` `calc` field (a `temp`+`calc` field is never
stored, so it always recomputes and creates no gap), so `to_canonical_bytes` and the
in-memory helpers `to_canonical`/`canonical_diff`/`is_canonical` are generated only then.
The verbatim/canonical choice can also be **carried on the value**: such a message gains a
wire-ignored `encode_mode` field (default `Verbatim`; set via the builder, `with_encode_mode`,
`set_encode_mode`). It is consulted by exactly one entry point — the `std`-writer
`encode(w)` — so you can set the policy once and stream the value; the explicit `to_bytes`/
`to_canonical_bytes` (and the `BitEncode` sink methods `bit_encode`/`canonical_bit_encode`)
ignore it. The field is **excluded from `PartialEq`/`Eq`/`Hash`/`Debug`** (a render preference, not
data — `#[bin]` intercepts those derives), which means such a type is constructed via the
builder, `new(fields…)`, or `decode` rather than a struct literal.
`validate = path` (the construction-soundness check `build()` runs) is also exposed as
re-runnable `validate()` / `is_valid()` methods: `build()` checks once, but a value can be
mutated before sending, so these re-check the *current* value (computed, never a stored
flag). By convention `validate` expresses **semantic** soundness — not the representational
`calc`/`reserved` fields — so validity holds for the canonical form too; `to_canonical_bytes`
stays a pure normalization (compose `validate()` before sending if you want the check).
## 6. The I/O ladder
For a buffer already in hand, the slice entry points decode in the message's *own* byte/bit
order: `decode_exact` (one message, reject trailing bytes), `decode_all`/`decode_iter` (every
message in a `&[u8]`), and `peek` (no consume). For a cursor — a stream, a socket, a growable
buffer — `decode(&mut Source)` is the single cursor decode (and `BitEncode::bit_encode` writes
into any `Sink`):
| `BitReader` | `&[u8]` | free (cursor math) | in-memory bytes |
| `StreamBitReader` | any `Read` | no (forward only) | a stream read once |
| `BufSource` | any `Read` | yes (bounded retain-and-seek) | a socket that also seeks |
| `BitBuf` | owned `Vec<u8>` (pushable) | yes (cursor math) | incremental framing: push bytes, pull messages |
| `SeekReader` | `Read + Seek` | yes (via `io::Seek`) | a large file/container |
| `BytesReader`/`Writer` (`bytes` feature) | owned `Bytes` | yes | zero-copy async framing |
A `Source` carries its own byte/bit order, so `decode(&mut Source)` reads in the *cursor's*
order — correct for the msb/big default, but a `little`/`lsb` message wants a cursor built with
its layout (`BitReader::with_layout(&v, Msg::LAYOUT)`). The slice entry points sidestep that by
baking `Msg::LAYOUT` in, so they are the foolproof "decode this buffer" path; `decode(&mut Source)`
is for streaming and custom sources.
**Byte order × bit order — the natural-layout rule.** The two knobs compose by one rule:
each bit order has a *natural* byte layout — the bytes the bit cursor produces with no
transform. MSB-first emits a value's high bits first, so its bytes land **big-endian**;
LSB-first emits low bits first, so its bytes land **little-endian** (value bit *k* goes to
stream bit *k*: exactly the CAN/DBC "Intel" layout, `raw |= v << start; frame =
raw.to_le_bytes()`). The declared byte order swaps a **byte-multiple** value only when it
*differs* from that natural layout; sub-byte widths are never byte-swapped. So the two
identity corners are the two real-world conventions — `big`+`msb` is network order,
`little`+`lsb` is DBC-Intel/SMB — and the mixed corners (`little`+`msb`, `big`+`lsb`) are
the deliberate swaps. Pinned against the embedded DBC reference formula in
`tests/bin_lsb_dbc.rs` (golden + property-tested), and golden at every layer in
`bin_order_matrix.rs` (message) and `bitstream.rs::cursor_layout_matrix` (cursor). The
`#[bitfield]` layer agrees by construction: `bits = lsb` packs `v << offset` into the
backing integer and `bytes = le` emits `to_le_bytes` — the same DBC layout.
Seeking is only needed by a message that uses `restore_position`; everything else runs
over the forward-only `StreamBitReader` too. **Seeking is a source capability, enforced
in the type system:** when a message uses `restore_position`, the generated
`decode` is bound on `SeekSource`, so decoding it through a forward-only stream is
a compile error rather than a runtime surprise; `forward_only` is the opt-in that
forbids seek directives outright. The in-memory cursor needs no `Seek` trait at all —
the whole buffer is in hand, so a seek is just cursor arithmetic (which also enables
e.g. DNS name-compression pointer following).
Above the cursor ladder, the opt-in **`net`** feature adds whole-message helpers over `std`
sockets — `MessageStream` (`read_message`/`write_message` over any `Read + Write`: one value drives
both directions of a `TcpStream`, no `try_clone`) and `MessageDatagram` (`send_message`/
`recv_message` over a sealed `DatagramSocket` — see §8). Both decode in the message's own layout,
and `MessageStream` reuses `BitBuf`'s framing rather than re-rolling it. The **`mock`** feature adds
in-memory `MockDatagramSocket`/`MockStream` (scripted inbound, captured outbound, chunked delivery,
error injection) so `net` code is unit-testable without a real socket.
### 6.1 `no_std` and the `std` feature (Option A)
`bnb` is `no_std` + `alloc`. `alloc` is unconditional — the codec's output model *is*
`Vec<u8>` (and `count` payloads / error strings own heap), so a heapless variant would
be a different crate, not a feature. The default-on **`std`** feature adds only the
rows of the table above that are backed by `std::io` (`StreamBitReader`/`BufSource`/
`SeekReader`, the `as_read`/`as_write` views), the `From<std::io::Error>` bridge +
`ErrorKind::Io`, and the `encode(writer)` convenience. The
forward-only/seekable distinction is unchanged; `no_std` simply has fewer `Source`
implementations to feed `decode` (the in-memory `BitReader` and `BitBuf`, and `BytesReader`
under `bytes`).
The chosen boundary is **buffer-at-a-time, not streaming** ("Option A"): `no_std`
decodes from a `&[u8]` and encodes to a `Vec<u8>`, then the caller writes those bytes to
its transport. This fits the workspace's datagram-oriented protocols (a UDP/ICMP/DNS
packet arrives whole) and keeps the change small and dependency-light. Two consequences
fall out of *a proc-macro cannot see the consumer crate's feature flags*:
- **`encode(writer)` is a blanket extension trait, not a generated inherent method.**
The single `EncodeExt` is `std`-gated and blanket-implemented over `BitEncode`, so it
appears exactly when `bnb/std` is on — whereas a `#[cfg(feature = "std")]` emitted into a
generated method would key off the *user crate's* feature name and silently vanish for a
default `cargo add bnb`. It dispatches to `bit_encode` vs `canonical_bit_encode` by the
value's [`encode_mode`](EncodeMode) — a settable, wire-ignored field that a `reserved`/`calc`
message carries (default `Verbatim`), rather than a call-time argument: set the policy once
(builder/`with_encode_mode`) and stream the value. Both the canonical path and `encode_mode`
are **defaulted methods on `BitEncode`** (no separate `CanonicalEncode` trait), so `encode`
works for every message — one without `reserved`/`calc` has no field and stays verbatim.
`#[bin]` intercepts `Debug`/`PartialEq`/`Hash` so the mode never affects equality (a render
preference, not data), which makes these types builder/`decode`-constructed. Cost: callers
bring the trait into scope (`use bnb::prelude::*`); the `to_bytes`/`to_canonical_bytes` `Vec`
encoders stay inherent and unconditional (sink-writing uses the `BitEncode` trait methods).
- **`BitEncode` carries `const LAYOUT`** so the blanket `encode` can build a correctly
ordered `BitWriter` without the per-type layout literal the old inherent method had.
- **`#[br(dbg)]` is `std`-only.** It emits a `tracing` event, and `tracing`'s default
features link `std`; the workspace dep can't be overridden per-member, so `tracing` is
an optional dep pulled in by `bnb`'s `std` feature. An embedded build uses its own
logger. **A future "Option B"** (an in-house `bnb::io` `Read`/`Write`/`Seek`
abstraction, à la `embedded-io`) would unify the code path and bring streaming to
`no_std`; it is deferred until an embedded byte-stream transport (TCP/serial) needs it.
## 7. Dual-use by default
The crates are **compliant by default, deliberately violatable**:
- **Builder defaults are compliant**, but the fields stay settable.
- **Parsers accept representable-but-non-compliant values** — unknowns are modeled as
data (`#[catch_all]`, retained flag bits), never hard errors.
- **Policy lives on the construction path, never in a parser.** `validate` gates
`build()`; decoding stays permissive, so hostile input can be parsed for analysis
but a malformed message can't be accidentally *built*.
- **Raw constructors never validate** (`from_raw`/`from_bits`, the `pub`-field struct
literal) — the open escape hatch.
Only the *physically unencodable* is refused (a value that doesn't fit its field's
bits), never the merely non-conformant. The one place a decode can panic is a `closed`
enum fed an out-of-set discriminant — which is exactly why `closed` is an explicit
opt-in and the default for untrusted input is `#[catch_all]`.
## 8. Key implementation decisions
- **The `temp` + `calc` anti-drift pattern.** A length/count you don't want to store is
read into a `#[br(temp)]` local and recomputed on write via `#[bw(calc = …)]`, so the
two directions are generated together and can never disagree with the data they
describe.
- **`validate` is construction-side only.** Auto-validating on decode would violate the
dual-use rule (never reject representable input), so a soundness check runs in
`build()` and surfaces as `BuilderError::Invalid`; the parser stays permissive.
- **No untrusted pre-allocation.** A `count`-driven `Vec` grows by pushing (each
element consumes ≥1 bit), so an attacker-controlled count can't trigger a giant
up-front allocation — it simply runs out of input and returns an error.
- **Context in two layers.** `ctx(...)` lowers to inherent `decode_with`/`encode_with`
(Layer 1 — covers nesting, counts, borrowed context with no `Args` type on the core
trait); a `DecodeWith<A>`/`EncodeWith<A>` companion (Layer 2) carries the same to
hand-written generics and trait objects.
- **One field-codec path — no `#[nested]` marker.** A field is either a `Bits` leaf (a
`uN`/`#[bitfield]`/enum — a single packed value, read by reading its bits) or a nested
message (another `#[bin]` type — a layout of fields, read by recursing into its codec).
A proc-macro can't tell the two apart by type name, so the codec once needed an explicit
`#[nested]` marker. Instead, **every `Bits` leaf also implements
`BitDecode`/`BitEncode`/`FixedBitLen`** (thin delegations to its bit read/write), so
`#[bin]` calls those uniformly for *every* field. These are **concrete** impls — one per
leaf type, emitted by each `Bits`-producing macro — *not* an `impl<T: Bits>` blanket,
which Rust's coherence rejects against the per-message derives (no specialization, no
negative bounds). The `Bits` *packing* role is untouched; only the stream-codec impls
were added. `#[nested]` is still accepted as a no-op for backward compatibility.
- **One cursor decode.** Decoding funnels through a single `decode(&mut Source)` over the I/O
ladder, plus the layout-baking slice helpers (`decode_exact`/`decode_all`/`decode_iter`/`peek`).
An earlier byte-cursor `decode(&mut &[u8])` was removed: a `&[u8]` cursor advances only whole
bytes, so it silently misframed messages that don't end on a byte boundary — the one thing a
*bit* codec must get right. `BitReader::new(&v)` is *also* zero-copy, so
`decode(&mut BitReader::new(&v))` replaces it at no cost with a bit-correct cursor. The price is
that a cursor carries its own byte/bit order (§6); the slice helpers bake the message's, keeping
the foolproof "decode this buffer" path foolproof.
- **Position-aware errors.** A codec error records the absolute **bit offset** where it
failed and the **field** being processed (the innermost wins, like a span), so a
failure points at the exact place. A streaming source that runs out mid-message
reports `Incomplete` ("read more and retry"), distinct from a definitive failure.
- **Reserved bits are explicit, stored, and observable.** A `#[reserved]` field is a
normal stored field with a known *spec value* (the type's zero, or the
`#[reserved_with(…)]` expression). On the verbatim path (`decode`/`to_bytes`) it
reads/writes its actual value, so a peer's non-compliant reserved bits are captured and
a caller can override them (dual-use); the builder defaults it to the spec value (so it
isn't required), and the **canonical** encoder (`to_canonical_bytes`) writes the spec
value instead. A *verified-on-read* constant is `magic` instead.
- **Sealed extension trait where the genericity is internal.** `DatagramSocket` (the `net`
feature) exists only to make `MessageDatagram` generic over `UdpSocket` and `UnixDatagram` — std
ships no datagram analog of `Read + Write` — not as an open extension point. So it is *sealed* (a
private `Sealed` supertrait): downstream can still *use* it as a bound (a generic handler) but
can't *implement* it. That keeps the 1.0 promise to two methods rather than a frozen,
openly-implementable trait, and leaves bnb free to evolve it — or to ship the feature-gated
`MockDatagramSocket` itself. Testing was the one real reason to implement it downstream, and the
`mock` feature covers that; a trybuild test locks the seal in. (`MessageStream` needs no such
trait — it's generic over std's `Read + Write`, so a `std::io::Cursor` or `MockStream` already
mocks it.)
## 9. Performance
Bitfields are plain shift/mask on a single backing integer — fully monomorphized, no
`bitvec`, no per-field heap, no runtime field tables. Benchmarked against the crates it
collapses (`bitbybit`, `modular-bitfield`) and a hand-written shift/mask baseline on an
identical DNS-shaped 16-bit field, `bnb` matches `bitbybit`, beats `modular-bitfield`,
and is within noise of hand-written (pack ~870 ps, unpack ~190 ps). The stream codec
takes a byte-aligned fast path — when a read/write is byte-aligned (the common case for
headers and `[u8; N]` payloads) it copies whole bytes instead of shifting one bit at a
time (~2–3× on aligned data); sub-byte reads fall through to the general bit loop. The
generated accessors and the runtime read/write methods are `#[inline]` so they inline
across crate boundaries.