Add ABI/annotation drift check; fix one dropped stakingV2 face

[taopunks]

getTotalColdkeyStakeOnSubnet(bytes32,uint256) is declared in precompiles/src/staking.rs with #[precompile::public] and #[precompile::view] inside impl StakingPrecompileV2, and dispatches on chain today. It is absent from precompiles/src/solidity/stakingV2.abi and stakingV2.sol.

Present in the Rust and missing from the vendored artifacts at both v438 (c1463f2c) and v440 (e4ffa2e1) — two releases. stakingV2.abi grew from 24 to 33 function entries across v439, so the artifacts are actively maintained and this looks like a drop rather than staleness.

An integrator enumerating callable faces from the vendored ABI — the natural thing to do, since it is the machine-readable artifact — cannot see this face at all.

Reproduce

Extract every #[precompile::public("…")] from precompiles/src/*.rs; extract every type: "function" signature from precompiles/src/solidity/*.abi; diff both directions. At e4ffa2e1: 168 rust faces, 168 abi functions, one present only in the Rust.

The reverse direction returns only verify(bytes32,bytes32,bytes32,bytes32) on ed25519Verify/sr25519Verify — those precompiles consume raw calldata and dispatch no selector, so the artifacts describe a calling convention with no annotated face behind them. Allowlisted in the check with that reasoning inline.

Two implementation notes

The Rust scanner is newline-agnostic on purpose. rustfmt wraps long attributes, so a line-oriented scan silently misses exactly the faces with the most parameters — today serveAxonTls and both long registerNetwork variants. A strict single-line regex finds 165 of 168.

The ABI side is module-agnostic, unioning every .abi rather than pairing each to a Rust module. Modules are snake_case, artifacts are camelCase, and staking.rs alone declares both staking and stakingV2 — a per-module mapping is a second thing that can be wrong, and it fails in the direction that manufactures findings. (My first attempt did exactly that and reported 9 false drops.)

Commits

1. the check — fails on main today, naming the drifted face
2. the fix — adds the entry to both artifacts; the check passes

Verification

Run against 1.89.0 (matching rust-toolchain.toml), with the workspace's expect-used / unwrap-used / indexing-slicing / arithmetic-side-effects deny lints replicated:

• cargo test --test abi_drift on pristine main → fails, naming getTotalColdkeyStakeOnSubnet(bytes32,uint256)
• same test with the fix applied → passes
• cargo clippy --all-targets -- -D warnings → clean

The test module carries #![allow(clippy::expect_used, clippy::arithmetic_side_effects, clippy::indexing_slicing)], matching the pattern already used in address_mapping.rs, alpha.rs, crowdloan.rs, leasing.rs and others. It uses no .unwrap().

No runtime logic is touched — requesting the no-spec-version-bump label.

Related

#2455 (precompile versioning and deprecation lifecycle) describes the general fragility this sits inside. #2998's coverage-and-testing.md specifies finding drift in both directions between the Rust signatures and the vendored ABIs — this adds a mechanical check for that specific direction, and the one face it currently finds is the motivation.

Disclosure

I hit this from the integrator side: I gate precompile calls from a contract and enumerate the callable surface to build that gate. The missing face is why I noticed.

[vercel]

@taopunks is attempting to deploy a commit to the RaoFoundation Team on Vercel.

A member of the Team first needs to authorize it.