ic402 v2.18.1

Patch — ic402 now builds with moc 2.1.0, with every dependency refreshed. Requested by EngramX. No
API, type, PaymentResult, stored-state or example.did change. The library source still compiles
with moc 1.x, so consumers move to moc 2 on their own schedule.

Changed

• Toolchain:

  • moc 1.16.1 → 2.1.0, wasmtime 48.0.5 → 50.0.0, and the mops CLI pinned in CI 3.4.2 →
    3.5.1. 3.5.1 closes smol-toml GHSA-r4xh-jqrq-34v2; 3.5.0 had already fixed mops publish on
    moc 2, which our release job avoids anyway with --no-bench.
  • Under moc 2.1.0 every module compiles with zero errors, and the full test suite passed
    unchanged.
  • [moc] args = ["-E", "M0074"] is gone: moc 2 makes M0074 an error by default.
  • Timers (moc 2.1.0, #6440). A timer dispatch rejected for lack of cycles no longer stops timers
    for good. ic402's recurring sweeps are unaffected, since an overdue run is still skipped. Its
    one-shot install timers (HMAC seed, EVM recipient derivation, content-store seed) are now retried
    instead of dropped. ic402 never cancels a one-shot, so the new limit that a retried one-shot can't
    be cancelled doesn't apply.

• Motoko dependencies:

  • sha2 0.2.5 → 0.3.2. 0.3 removed Digest.fold/combineLeaves/combineNodes and changed the
    stored Sha256.Digest type. ic402 only uses the one-shot fromArray/fromBlob, so its hashes are
    byte-identical and every golden vector (escrow subaccounts, Ed25519, vouchers) passes.
  • ecdsa stays at 8.0.1, on purpose. 8.0.2 only migrates it to moc 2 and moves it to sha2
    0.3, but it declares that as sha2@0. mops resolves one version per 0.x major, so 8.0.2 drags
    every package in a consumer's build that asks for any 0.x sha2 onto 0.3. Packages written
    against 0.1/0.2 then fail to compile: the EngramX canary caught it on ic-certification 1.1.0,
    whose MerkleTree.mo calls the SHA256.Digest(#sha256) constructor that 0.3 removed. ic402's
    own sha2 = "0.3.2" has no such effect, since it uses the unversioned package name. 8.0.1
    compiles and passes under moc 2.1.0.
  • core 2.6.2 comes in transitively. base stays at 0.16.0, its final release.

• Published npm dependencies:

  • @icp-sdk/core 6.1.1. It bounds Candid decode allocation and LEB128 parsing against
    memory/CPU-exhaustion DoS (#1393, #1391), and reuses the signed call on v4 fallback (#1405).
  • @modelcontextprotocol/sdk 1.32.2, viem 2.57.13, cborg 6.1.6, zod 4.6.7.

• Dev tooling:

  • vitest 4.1.11 → 5.0.5. A major. Two of its breaking changes touch this repo:
    • Config files are no longer found in parent directories. packages/client now has its own
      vitest.config.ts, so pnpm test:client keeps the root config's 120 s timeout.
    • clearMocks is now on by default. That's a no-op here, since every mock is created inside its
      test.
    • Browser mode, benchmarks, coverage and nested vi.mock aren't used.
  • eslint 10.13.0, typescript-eslint 8.72.1, prettier 3.9.10, @types/node 26.6.5.
  • icp CLI 1.7.0, as both the devDependency and the CI pin. It only adds commands, none of which
    change the ones ic402 runs; checked locally as the pin note requires.
  • pnpm 10.34.9.

• Fewer outstanding calls on timer and install paths. A local async function called with
  await is a self-call, and on a 13-node subnet each one holds about 82B cycles of reply
  reservation until it answers. These now run inline with await*. Behaviour is identical for any
  ICRC-1-conformant ledger. The only difference: a ledger reply that fails Candid decoding now rests an
  ICP close in #closing for forceResolveSession, as a malformed icrc1_fee reply already did,
  instead of reopening the session or recording an owed refund.

  Path	Self-calls saved
  Install-time EVM recipient derivation (the startTimers one-shot)	2 → 0
  deriveEvmRecipient, when a consumer calls it (its own async hop remains)	2 → 1
  Escrow settle and refund inside an ICP close, from the sweep or the payer	2 per close
  recoverEscrow, and openSession's concurrent-limit refund	1 each
  ServiceRegistry's payment from the main account, for job payouts and the job sweep's refunds	1 each

  • Escrow.settle/refund keep their async signatures and gain inline settleCore/refundCore.
  • Deliberately still separate messages: the sweep's per-session close (2.17.3: a trap fails only
    that close) and refundOnRail (it parses the buyer principal before its first await, so one
    malformed job can't wedge the job sweep).

• Removed 6 unused identifiers in the library (moc 2 warns M0194).

Interop note: viem now refuses colon-namespaced EIP-712 names

• What changed: viem ≥ 2.57.12 (wevm/viem#5308) rejects struct names such as Hyperliquid's
  HyperliquidTransaction:Withdraw with InvalidStructTypeError. viem ≤ 2.57.11 accepted them.
• Effect: ic402's encoder is unchanged and still signs them, as 2.13.1 intends. A verifier on
  current viem can't hash that typed data; ethers, and viem ≤ 2.57.11 on the venue's single-struct map,
  still agree with ic402. @ic402/client never hashes typed data with viem, so it isn't affected.
• Test: packages/client/test/eip712-colon-interop.test.ts is re-derived. It now checks the Motoko
  golden digest against an independent EIP-712 computation (keccak + ABI encoding). A control shows
  that computation equals viem's on a plain name, and the drift alarm pins viem's refusal.
• Unchanged: the EIP712Domain: first-segment rejection stays (fail-closed).

Not changed

• TypeScript stays at 6.0.3. typescript-eslint 8.72.1, the latest, still requires TypeScript
  < 6.1.0, so TypeScript 7 is blocked by that dependency.
• pnpm 11/12 stay deferred, as they're package-manager majors. The move from base to core
  belongs to 3.0.0.
• The example keeps persistent actor and preupgrade/postupgrade, so it still compiles under
  moc 1.x. moc 2 only warns (M0217, M0270).