Files
zeekay ca4572759c Every Rust chain reads and writes the wire a schema states
The P-chain, the Z-chain and the F-chain join the Q and X chains: what each
puts on a wire is stated once in chains/schema, and zapgen writes the
accessors. Five schemas, five emitted modules, and one runtime under all of
them — chains/rust/zap, also emitted. The five hand-written copies of that
runtime are gone: 4,056 lines that were the same reader written out five
times, three of them byte-identical and two variants nobody had reconciled.

WHAT THE VARIANTS DISAGREED ABOUT, and it was not cosmetic. Every hand copy
defaulted `Builder::new` to wire version 2; the emitted runtime keeps `new`
at version 1 and names `new_v2` for the other. So the swap silently moved
byte 4 of every message the P-chain wrote outside a generated builder, and
the warp differential caught it against the bytes Go wrote — a one-byte
divergence in a header nobody reads by hand. Every P-chain writer now says
new_v2, which is what it always wrote.

THE SCHEMA WAS PROVED BEFORE THE TABLE WAS DELETED. A test asserted every
offset and every size in pchain.zap against the constant the chain had been
using — 27 structs, every field — and it passed before a line was removed.
What replaces it now compares schema to schema: every kind's envelope is
Base's envelope, checked field by field against the emitted constants, so
`read_envelope` reading any transaction through Base's accessors is a thing
that is checked rather than a convention to remember.

THE BYTES DID NOT MOVE, and every chain says so against the corpus the Go
chain generated: 68 P transactions and 4 P blocks, 19 X transactions and 6 X
blocks, 12 Q blocks, 22 Z blocks, 43 F transactions, each re-encoded from the
values that were read and compared to Go's hex. Those are every vector the
corpus expects to parse; the rest are the deliberately malformed ones, and
they are still refused. `make chains` is 725 vectors, 4 implementations, 13
evaluators, AGREED on every compared field, exit 0, nothing under NOT
ANSWERED.

Two fields of one name and a struct named but not declared are now refused by
zapgen's front end, once, for every backend. The P-chain hit the first:
an envelope's owner-address POOL and the owner group a transaction hands
ownership TO were both called OwnerAddrs, and the emitted Rust had the same
struct member twice. They are named apart now — NewOwner on a chain
ownership transfer, NetworkOwner on a network being created.

WHAT IS STILL HAND-WRITTEN, and it is the P-chain only: warp.rs and
warpmsg.rs (the warp envelope and its payloads), genesis.rs, persist.rs (the
disk state), and components.rs (the fx primitives it spends). Those are real
remaining offsets and they are next.

The fx primitives belong in neither chain's schema — the X-chain makes them
and the P-chain spends them, so xchain.zap and a P-chain statement of the
same TransferOutput would be two statements of one format. Lifting them into
a shared fx.zap needs an import across schemas that zapgen does not have; the
attempt is what found the undeclared-struct hole, and the refusal now says so
at the schema rather than in emitted Rust.

378 platformvm, 290 xvm, 110 quantumvm, 150 zkvm and 97 fhevm unit tests, the
five corpus byte tests, the golden vectors and the differentials all pass.
2026-09-06 13:54:26 -07:00

55 lines
1.9 KiB
Rust

// SPDX-License-Identifier: BSD-3-Clause-Eco
// Copyright (C) 2019-2026, Lux Industries Inc. All rights reserved.
//! Every Z vector in the shared corpus, written back byte for byte.
//!
//! The corpus was generated by the Go chain, so a vector's hex is what Go
//! wrote. Reading one and writing it again has to reproduce it exactly: this
//! chain's writer and Go's are the same writer or they are not. That is what
//! makes the corpus a check on the WRITE path and not only on the read path —
//! a reader can agree about every field of bytes it could never have produced.
use std::path::PathBuf;
use lux_zkvm::wire;
fn corpus() -> String {
let p = PathBuf::from(env!("CARGO_MANIFEST_DIR"))
.join("../../../conformance/corpus/vectors.tsv");
std::fs::read_to_string(&p).unwrap_or_else(|e| panic!("{}: {e}", p.display()))
}
fn unhex(s: &str) -> Option<Vec<u8>> {
if s == "-" || s.len() % 2 != 0 {
return None;
}
(0..s.len() / 2)
.map(|i| u8::from_str_radix(&s[i * 2..i * 2 + 2], 16).ok())
.collect()
}
#[test]
fn every_z_vector_is_written_back_exactly_as_it_arrived() {
let text = corpus();
let mut blocks = 0;
for line in text.lines() {
if line.is_empty() || line.starts_with('#') {
continue;
}
let f: Vec<&str> = line.split('\t').collect();
if f.len() < 5 || f[0] != "V" || f[2] != "Z" || f[3] != "block" {
continue;
}
let id = f[1];
let Some(bytes) = unhex(f[4]) else { continue };
let Ok(body) = wire::read_block(&bytes) else {
continue;
};
assert_eq!(wire::write_block(&body), bytes, "{id}: block bytes moved");
blocks += 1;
}
// A corpus that stopped naming Z vectors would pass every assertion above.
assert!(blocks > 0, "no Z block vector was read back");
println!("{blocks} Z blocks written back byte for byte");
}