feat(wire)!: standardize the DIG peer wire on native types, drop the chia bridge

[MichaelTaylor3d]

DO NOT MERGE — gate round not yet run. Publishes to crates.io before dig-gossip can consume it (§3.6, release-first).

Standardizes the DIG peer wire on native DIG types and removes the on-ramps back to the forked route.

Blast radius checked

Every chia* reference in every src/ and tests/ file, enumerated by grep on this tree, plus a
read-only measurement of the downstream consumer (dig-gossip) via git show origin/main:<path>.
gitnexus index not built for this worktree — fell back to grep + direct read per §2.0 bound 2.
Risk: HIGH — this is a live network's wire format and a breaking API change to every consumer.

Byte-identity: proven, not asserted

Commit 1 pins 9 absolute golden vectors captured on the pre-change tree, as their own commit,
before a single type was touched. Commit 2 makes the change. All 9 pass unchanged.

The vectors are load-bearing: flipping the DigMessage id length prefix to little-endian fails two
of them. They exist because tests/wire_compatibility.rs is a relative proof — it compares
DigMessage against chia_protocol::Message, so an upstream layout change moves both sides
together and it stays green while the network breaks.

No wire change. No coordination event. Shape breaks only.

What changed

• DIG-owned Bytes and NodeType. Encodings deliberately identical to what they replace: a
  u32 big-endian length prefix, and the frozen 1..=7 role discriminants. A byte naming no role
  is refused, never defaulted — a zero byte is what a truncated buffer produces.
• The four chia_protocol::Message bridge helpers are deleted. They were the on-ramp back to
  the fork: a consumer that could cheaply obtain a Message kept building one, and the closed
  ProtocolMessageTypes enum came back with it.
• The blanket pub use chia_protocol::* is replaced by three named re-exports.

Deliberately KEPT, with the reasoning recorded in lib.rs

Recorded in the crate so nobody "finishes the decoupling" later by removing a dependency that was
assessed and retained:

• chia-traits + chia_streamable_macro — a trait and a derive macro. Neither has the property
  this crate escapes: no closed enum, no private-field wire authority. Replacing them means owning a
  serializer — new surface and a fresh byte-identity risk — to buy nothing.
• ChiaProtocolMessage, TimestampedPeerInfo, chia-sdk-client — genuine chia traffic. A DIG
  node talks to chia full nodes too. Chia types for chia traffic is the design.

The rate-limit hazard, now tested

rate_limit.rs's header has always warned that a chia-protocol/chia-sdk-client version split
silently collapses every Chia opcode to default_settings — a large loosening, no compile
error, no panic. Nothing tested it. A silently permissive rate limiter is a DoS surface.

Three tests added. The pin uses absolute literals (opcode byte + both limit numbers), not a
comparison against V2_RATE_LIMITS: that would ask the same possibly-shifted enum for its own key
and agree with itself. Backed by a tighter-than-default guard (so a collapsed table is
distinguishable from a re-keyed one) and an entry-count floor.

The absolute pin immediately caught a transcription error of my own — TransactionAck is in the
tx map, not other.

Version

0.4.0. Breaking on a 0.x line is a minor bump. main is 0.3.0, the only other branch
(feat/2228-dig-peer-link) also claims 0.3.0, and no tag claims 0.4.0 — no collision.

Gates

cargo fmt --check, cargo clippy --all-targets --all-features -D warnings, and
cargo test --all-features all exit 0, asserted by exit code rather than piped output.
86 tests pass (68 lib + 9 golden + 5 inbound + 4 liveness).