buffa-types: back Any.value with bytes::Bytes for refcount-bump clone

[kollektiv]

google.protobuf.Any is commonly cached and cloned into repeated google.protobuf.Any response fields (RPC servers, fan-out, message buses). With value: Vec<u8> each Any.clone() deep-copies the encoded payload; with bytes::Bytes it is one atomic refcount increment, regardless of payload size.

The change is driven from gen_wkt_types.rs (bytes_fields = [".google.protobuf.Any.value"]) so task gen-wkt-types and the check-generated-code CI job stay green. any_ext.rs helpers (pack, unpack_*, JSON/textproto paths) are updated; signatures are unchanged.

Benchmark

New buffa-types/benches/any_clone.rs:

Any.value.len()	Vec<u8>	Bytes	speedup
64 B	32 ns	34 ns	1.0x
1 KiB	43 ns	34 ns	1.3x
16 KiB	189 ns	35 ns	5.4x
256 KiB	5.99 µs	35 ns	171x
1000×1KiB into Vec<Any>	143 µs	54 µs	2.6x

Any.clone() is now constant-time in payload size; the ~34 ns floor is the type_url: String clone.

Compatibility

Wire format: fully compatible. Bytes derefs to &[u8]; encode (encode_bytes), decode (Bytes::from(decode_bytes(buf)?)), JSON (base64 of &[u8]), and textproto all produce/accept identical bytes.

Rust API: source-breaking (public field type changed). Common caller patterns:

Pattern	Result	Fix
Any { value: vec![..], .. }	fails	vec![..].into()
Any { value: Vec::new(), .. }	fails	Default::default()
let v: Vec<u8> = any.value	fails	any.value.into()
any.value.as_slice()	fails	&any.value[..] or .as_ref()
&any.value (deref to &[u8])	works
any.value.len() / indexing / is_empty()	works
Any::pack / unpack_if / unpack_unchecked	works
any.clone()	works (and now constant-time)

The .into() / Default::default() shims compile against both the current and the changed type (Vec<u8>.into() is identity for Vec<u8>, From<Vec<u8>> for Bytes), so callers can pre-adopt before this lands.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[kollektiv]

I have read the CLA Document and I hereby sign the CLA

[kollektiv]

recheck

[iainmcgin]

[claude code] Reviewed and pushed two follow-ups directly to your branch (maintainerCanModify):

• CHANGELOG entry noting the Any.value: Vec<u8> → Bytes break. Since 0.x crates treat any source-break as a minor bump, this likely means the next workspace release cuts 0.4.0 — flagging for release-management when you merge.
• clone_shares_payload_buffer unit test in any_ext.rs. Asserts as_ptr() equality after Any::clone(), pinning down the headline invariant of this change so a future refactor that accidentally reintroduces a copy (e.g. swapping .into() for Bytes::copy_from_slice) fails loudly instead of silently regressing the perf win. The bench covers speed; this covers "how it's fast".

Two perf follow-ups I spotted but didn't fix in this PR (they're out of scope / larger):

• AnyView::to_owned_message copies Bytes payload unnecessarily #52 — AnyView::to_owned_message still memcpys the payload via Bytes::copy_from_slice. Getting zero-copy here needs threading a Bytes parent through decode_view (codegen-wide change).
• decode_bytes could produce Bytes zero-copy via Buf::copy_to_bytes #53 — decode_bytes in merge_field allocates a Vec<u8> and wraps. Buf::copy_to_bytes would be zero-copy on Bytes-backed buffers.

Minor nits not worth blocking on (feel free to address in this PR or skip):

• bench_clone_into_vec uses a fixed "1024B" group label for both rows — harmless id collision.
• Consider a brief doc note on Any in any_ext.rs showing Bytes-first construction patterns now that .into() is ubiquitous at call sites.

Otherwise LGTM — clean use of the existing bytes_fields knob, correct clear() reassignment, correct last-wins merge semantics, and encode_to_bytes is the right constructor hook. Nice find.

[kollektiv]

recheck

[iainmcgin]

recheck

[iainmcgin]

[claude code] LGTM — clean use of the existing bytes_fields knob, correctness verified, no_std clean on host + bare-metal ARM, all CI green.