[predge-ai]

Why

For Arc Demo Day (16 Nov) we want to show Predge as the arbiter of Circle's Refund Protocol (circlefin/refund-protocol). PredgeRefundArbiterV2 was built against our MockRefundProtocolV2. To find out whether it works with Circle's real contract, we read Circle's public src/RefundProtocol.sol. No Circle code was cloned, installed or run.

Gap analysis: does the deployed arbiter work with Circle's contract as it is? No. It needs a small adapter.

What PredgeRefundArbiterV2 calls	Circle RefundProtocol	Gap
paymentInfo(id) -> (payer, payee, amount, releaseTimestamp)	does not exist; the public getter is payments(id) -> (to, amount, releaseTimestamp, refundTo, withdrawnAmount, refunded)	Blocking. Every fileEvidence and rule call would revert.
payer	Circle does not store who called pay; refunds go to refundTo, and refundTo can be changed with updateRefundTo	map payer to the current refundTo
amount, in native wei (sizes the bond)	in USDC's 6 decimals (Arc native gas uses 18)	scale by 1e12 on Arc
refundByArbiter(id)	same signature, onlyArbiter; reverts on already refunded or InsufficientFunds; no check on release time	compatible. The adapter turns "already refunded" into a no-op
setLockupSeconds(r, s)	same signature, onlyArbiter, cap is 180 days	compatible. Predge allows up to 365 days, so Circle reverts between 180 and 365 days
caller identity	arbiter is set once in Circle's constructor and has no setter	the adapter becomes Circle's arbiter and accepts writes only from the bound Predge arbiter
payment ids	start at 0	fine (the Predge status for id 0 is None)
unused by Predge	earlyWithdrawByArbiter, deposit/withdrawArbiterFunds	not exposed. A refund comes only from the recipient's balance in Circle; if that balance is short, finalize reverts and expire voids the ruling

The live Arc/Base PredgeRefundArbiterV2 points immutably at MockRefundProtocolV2. Pointing at Circle's contract therefore needs a new arbiter instance with refundProtocol = adapter. The bytecode does not change.

What this PR adds

• contracts/v2/interfaces/ICircleRefundProtocol.sol: our own interface, signatures only.
• contracts/v2/adapters/CircleRefundProtocolAdapter.sol: about 120 lines, no owner, holds no funds. A one-time bind checks the wiring in both directions.
• contracts/v2/mocks/MockCircleRefundProtocol.sol and MockUSDC.sol: our own local stand-ins that mirror Circle's interface and behaviour, for tests and demos only.
• test/foundry-v2/CircleRefundAdapter.t.sol: 14 tests covering release, refund on verdict, dispute (overturned to refund, refund upheld), the seller refunding first, already-refunded payments, zero lockup refused, the 180-day cap, a refundTo change after the ruling, and a 6-line demo log.
• scripts/foundry/DeployCircleRefundDemo.s.sol: deploy script for Arc testnet. Simulated only, nothing broadcast.

No change to PredgeRefundArbiterV2, config/set.v2.json or the deployed addresses.

Tests

FOUNDRY_CONFIG=foundry-v2.toml forge test: 166/166 pass (14 new). The node test plan-v2.test.mjs #2 ("would collide with a v2.0 address") already fails on v2/contracts without this PR and is unrelated.

Before mainnet

The adapter and both mocks are unaudited. Circle's RefundProtocol is itself marked unaudited and educational. Do not merge this until it has been reviewed.