[avrabe]

Refs #1649

This moves the wasm-tools family from 0.262 to 0.263: wasmparser, wasm-encoder, wit-parser and wit-component go to 0.263, wat to 1.263, and wast to 263.0. The bump is a cargo update -p of those crates plus wasm-metadata. The resulting Cargo.lock delta is the same as dependabot's. One change goes beyond #1649: the wat requirement is raised to 1.263, which dependabot left at 1.262. The lock resolves 1.263.0 either way.

API breaks (both in synth-core)

Site	Break	Fix
arena_bind.rs (allocator body)	wasm_encoder::MemArg replaced memory_index with target: MemArgTarget and added ordering	target: MemArgTarget::Memory(0), ordering: None. This encodes the same bytes as memory_index: 0.
wasm_decoder.rs (memarg helper)	wasmparser::MemArg replaced memory: u32 with target: MemArgTarget::{Memory(u32), Array(u32)}	The helper is renamed memarg_target. A Memory(k > 0) target is wrapped in MultiMemory as before. An Array target (a GC array payload) loud-skips the function instead of being lowered as a linear-memory access.

A semantic change that this PR switches off

Parser::new enables WasmFeatures::all(). In 0.263 that set gained two in-flight proposals, acquire-release-atomics and multibyte-array-access. Both give new meanings to memarg flag bits 4 and 5, which 0.262, and the standard wasm 3.0 encoding, read as part of the alignment exponent.

Six hand-built probe modules showed the bumped decoder diverging from base 617e440:

• Flags 0x12 and 0x32 became a hard Failed to decode error.
• Flags 0x20 and 0x22 consumed an extra LEB as an array type index, then loud-skipped the function.

The decoder now parses through decode_parser(), which removes those two features. All six probes are byte-identical to base again. Synth lowers neither proposal. The Array arm above is therefore unreachable today and stays only as a defence.

Two synth-core unit tests, memarg_bit{4,5}_is_alignment_*_1649, pin the 0.262 reading of flags 0x12 and 0x22. Paired control: with decode_parser() reverted to the default feature set, both tests fail.

These probes are invalid wasm (an alignment larger than natural), and base accepts them. That is the existing invalid-accept class, and this PR leaves it unchanged.

Evidence: base 617e440 vs this branch, both built with default features (riscv)

• Byte-identity census. 554 inputs (tests/**, scripts/repro, examples, plus the 257 top-level spec .wast files) × 8 configs give 4432 pairs. The configs are cortex-m4, cortex-m7dp, cortex-m4 --relocatable, the same with --debug-line, cortex-m4 --safety-bounds software, cortex-r5, riscv rv32imac and aarch64.
  • 1792 pairs produce byte-identical objects.
  • 2516 pairs are refusals with the same rc and the same diagnostics.
  • 124 pairs have identical rc and output files but different diagnostic text, in 16 simd_*.wast files. In each case the change is MemArg's Debug output inside an unsupported-op warning: memory: 0 } became target: Memory(0), ordering: None }. After applying that one substitution, 0 differences remain. spec_compile_census.py pins all 257 spec files and passes on the branch.
  • The census was rerun with --allow-skipped-exports, so a module with a declined export still emits an object. That run has 2208 byte-identical objects, 2100 identical refusals, and the same 124 pairs that differ in text only.
• Validation-acceptance parity. The census gives the same rc on all 4432 pairs. The six memarg-flag probes also match base in rc and output sha256.
• dissolve: bind passed-through embedder import cabi-arena-realloc to the native TCB arena allocator #418 arena-bind. allocator_body is the only site whose emitted bytes this change touches. cabi_arena_bind_418_differential.py passes on base and on the branch with identical logs. Its compile output is byte-identical: the ELF sha256 matches.
• Spec execution oracle (scripts/repro/spec_exec_1636_differential.py). It passes on both binaries: 35676 directives executed, 11634 on cortex-m4 and 24042 on cortex-m7dp. There are 53 failures, all exact baseline entries. The per-directive --json ledgers of base and branch are byte-identical.
• Builds and tests.
  • cargo fmt --check and cargo clippy --workspace --all-targets -- -D warnings pass.
  • cargo test --workspace --no-fail-fast: 3271 passed, 0 failed. That run predates the two new tests; after adding them, the synth-core lib tests pass 253 of 253.
  • --features riscv (async_intrinsics_gate): 2 passed. -p synth-cli --features verify: 479 passed. -p synth-verify --features z3-solver,arm: 278 passed.
• Claim Check. All 55 steps of the Claim Check job, run locally: 0 failures.
• Pins. scripts/check_version_pins.py passes. Bazel resolves crates from Cargo.lock (crate.from_cargo), so neither MODULE.bazel nor crates/BUILD.bazel pins these crates. MODULE.bazel.lock is not kept in sync with crate bumps: the 0.262 bump (chore(deps): bump the wasm-tools family to 0.262 #1591) did not touch it either, so this PR leaves it alone.

Rebased onto 5dc4e28

The evidence above was measured on base 617e440. On base 60170da (after #1697, #1702 and #1705) the branch was re-checked: cargo fmt --check, clippy clean, cargo test --workspace --no-fail-fast 3299 passed / 0 failed, Claim Check 55/55, and the spec-suite workflow steps (compile census PASS on all five backends, #1225 PASS, #1441 PASS, spec execution #1636 PASS: 35715 directives, baseline 54/54).

It was then rebased onto 5dc4e28, which adds only [profile.dev.package.ordeal*] to Cargo.toml (#1706) in a hunk disjoint from this one. On that head only cargo metadata --locked and cargo fmt --check were run locally, because the host was out of disk; CI is the full check for this head.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L

[codecov]

Codecov Report

❌ Patch coverage is 89.23077% with 7 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
crates/synth-core/src/wasm_decoder.rs	88.88%	7 Missing ⚠️

📢 Thoughts on this report? Let us know!