Implemented PRIO, six application contracts, ABI exports, vendored dependencies, operator tooling, and documentation.

Verified: forge build, 39 Foundry tests, forge fmt --check, and four operator tests—all pass.

Deployment requirements and remaining release gates are documented in README.md, including hook-capable factory support and job.open availability. No transactions were broadcast.

Created launch.json with PRIO and six application contracts in dependency order.

Schema and ABI checks passed. forge build passed with existing lint warnings; all 39 tests passed.

Only launch.json changed. Its notes preserve the unresolved hook/factory compatibility finding.

The review is complete. .imd-findings.json at the repository root holds four findings and a coverage row for all 51 listed entry points. No source, test, config or doc file was changed; the one scratch test I used was deleted after it confirmed a finding.

What I covered. The Execution Trace, Periphery and First Principles guides, applied to all seven contracts and the SwapPreview library. I rebuilt the project and ran the existing suite (39 tests pass), compared the hook's delta sign conventions line by line against the pinned v4-core Hooks and Pool libraries, and confirmed every file in docs/abi is byte-identical to the compiled ABI. Arena, PrioStaking, RevenueBuyer and the hook's swap-fee math held under every failure mode in my guides. The Arena oracle timing is sound: an answer cannot land before revealEnd or after the shared deadline, so the cancel and settle paths are mutually exclusive.

Findings, most severe first.

• Medium, TreasuryFeeHook.sol:67. The hook needs a salt-mined address with permission bits 0x20cc, an owner-only configure call before pool initialization, and a pool key naming it as the hook. A one-transaction generic factory that initializes with PoolInitializationGuard satisfies none of these, so the launch either reverts or opens an unhooked pool. The author already flags this in docs/REVIEW.md. It is the fact that decides whether this launch can deploy, and the manifest reviewer needs the univ4_hook factory's actual lifecycle as evidence.
• Low, OracleAdapter.sol:89. The adapter rejects any attestation whose issuedAt is ahead of block.timestamp, though the protocol base contract allows five minutes of clock drift. Verified with a scratch test: one second ahead reverts, the same signed attestation is accepted a second later. On the paid callback path there is no retry, so the 0.5 IMD is spent and the question stays pending until someone resubmits manually.
• Low, FeeTreasury.sol:191. Buying PRIO rewards requires the IMD and Intake services to be configured although it uses neither. On a chain without an Intake contract, the owner cannot call setServices honestly, and the 30% reward bucket is unreachable.
• Info, SwapPreview.sol:63. The 128-step quote cap means a swap crossing more than 128 initialized ticks reverts. Dust positions can raise that count, but it only bounds swap size, not the pool, and the author documents it.

Static-analysis leads. The reentrancy-balance lead on buyTokens is the hook paying a fee back to the treasury during its own purchase. I traced it and the balance-versus-accounted invariant holds, so I did not report it. The strict-equality and divide-before-multiply leads in PrioStaking are benign by construction.

Not reached. No live factory, router or chain fork was available, so the deployment finding rests on the task's description of the factory and the author's own notes rather than an executed launch.

The review is complete. The findings file is written, every snippet matches its source line, the scratch directory is removed, and the working tree is otherwise untouched (only the untracked .imd-findings.json remains, which git ignores).

Outcome. The math in this project is sound. I traced the hook's 0.5% fee in all four swap modes, the Synthetix-style staking accumulator, the treasury's reserve-then-30/30/40 split and PRIO halving, the arena's 102-PRIO escrow settlement, and the oracle adapter's windows against the Math Precision, Boundary and Numerical Gap checklists. No loss-of-funds or wrong-recipient defect exists. The three findings in .imd-findings.json are:

• Medium, deployment path. TreasuryFeeHook.beforeInitialize reverts NotConfigured until the owner has called configure, so a factory that deploys and initializes the pool in one transaction cannot launch this hook. The code fails closed as the brief asks, but no accepted launch path currently satisfies it. The builder already flags this in the README and review handoff.
• Low, fee rounding. The exact-input ETH buy path iterates to an integer fee fixed point and reverts FeeRoundingConflict when none exists. Measured on the test pool, 14 of 3,000 consecutive amounts starting at 1 ETH revert, including 1000000000000000176 wei, while the 1-wei neighbour succeeds. The quote reverts too, so the website cannot display a fee for those amounts. Accepting the larger of the two oscillating values would fix it within 1 wei.
• Low, timestamp boundary. OracleAdapter._submit rejects an attestation whose issuedAt is one second ahead of the block timestamp, stricter than the protocol's five-minute tolerance. On the paid callback path that single delivery attempt is lost and must be re-submitted manually later. Verified: the same signed attestation is accepted one second later.

Coverage. All 52 listed entry points have a row, plus four invariant rows. Three are marked finding, the rest holds. I also confirmed that every exported ABI in docs/abi is identical to the compiled artifact, and the full existing suite of 39 tests passes. Static-analysis leads I checked and dismissed with evidence: the reentrancy on buyTokens is the hook's recordFee re-entering during the purchase, and the accounting stays consistent; the divide-before-multiply in startRewards is deliberate, with the remainder kept queued.

The review is complete. The findings file is at the repository root, and the harness's own exclude list hides it and the scratch tests from git, so no tracked file changed.

Outcome. The permission model is sound: every privileged write is owner-gated, every callback checks its caller, all constructors take $owner rather than msg.sender, and no admin path reaches player escrow, staking principal or reserved prizes. The ABI exports match the compiled artifacts for all seven contracts. The one blocking item is a launch-path conflict, not a permission bypass.

Findings written to .imd-findings.json (five, each with a confirmed reproduction under test/scratch/):

1. Medium, launch-path conflict. The hook can only be activated after an owner-only configure call and only by the configured initializer. Under the supplied one-transaction ProjectFactory flow that call cannot happen before the factory initializes the pool, so a factory initialize reverts, an unmined CREATE2 salt aborts the launch at the hook constructor, and a guard-hooked pool charges no 0.5% fee at all. The scratch test shows all three outcomes. The manifest step cannot be accepted until a hook-capable path or a constructor-configured hook exists.
2. Low, trust gap. The treasury pays for an oracle callback request without checking that the adapter's delivery address is the Intake. With delivery unset, the paid callback reverts, is never retried, the IMD is spent, and delivery cannot be fixed while the question is pending.
3. Low, asymmetry. The adapter rejects attestations whose issue time is one second ahead of the block, while the inherited base allows five minutes of clock drift. A drifted paid callback is dropped.
4. Info, trust assumption. The owner can route all new revenue to the operator wallet through an unbounded reserve target, bypassing the 30/30/40 split. The README's claim that ownership transfer cannot redirect the paying wallet's share holds only for ETH already allocated.
5. Info. All five privileged contracts use single-step ownership with renounce enabled, so one wrong call permanently disables rounds, questions and configuration.

Coverage. All 52 listed entry points have a row, plus three invariant rows and one honest unreached row for things I could not exercise locally: the live factory lifecycle, the chain's actual LaunchFees fee, and callback gas when the signer is an ERC-1271 registry.

Leads checked and dropped. The Slither reentrancy lead on buyTokens is benign by construction: the only reentrant write is the hook crediting fresh unallocated revenue. Permissionless startRewards can delay later funding by at most one period and cannot exclude it. The v4 self-sender hook skip is unreachable because the hook never swaps.

Added failure-path tests and Arena/treasury invariants; strengthened staking invariants. All changes are under test/.

• forge build passed.
• Offline forge test: 68 passed, 0 failed, 0 skipped.
• Invariants exercised 65,536 randomized calls.

No confirmed implementation defects found. Live-chain integration remains unverified.

The review is complete. .imd-findings.json is written at the repository root with 4 findings and 57 coverage rows, all four snippets verified against the tree, the scratch probes deleted, and the untouched suite still passing (39 tests).

Outcome. The economics of this project are sound. I found no critical, high or medium defect in the Economic Security, Invariant or Flow Gap areas. The four recorded items are one low-severity liveness defect and three informational notes.

What I verified in depth

• Arena conservation. I fuzzed 400 seeds of four consecutive rounds across all three modes with partial reveals, ties, boss-threshold misses and oracle-timeout cancels. The token balance always equalled free rewards plus reserved prizes plus escrow liability, and after all claims both escrow and reserved balances were exactly zero. Fees and penalties flow to future prizes as the brief requires.
• Treasury ETH accounting. Every writer of the five buckets keeps accountedETH equal to their sum, including the self-fee re-entry when the treasury buys PRIO through its own hooked pool.
• Staking backing. The stateful invariant suite holds. Empty-stake time is recycled, dust stays in reserve, principal is never a reward source.
• Hook fee. All four swap directions charge floor(nativeBase/200) and deliver the whole fee to the treasury. I checked the exact-output gross-up algebra by hand.
• Static-analysis leads. The reentrancy-balance, divide-before-multiply, locks-ether and unused-return lines all trace to safe or intended behaviour. None reproduced as a defect.
• Killed lead. The paid oracle callback fits the 200k stipend with headroom. With cold storage, a binary search found the minimal gas at about 165k for an EOA signer and 173k for a three-slot ERC-1271 registry.

Findings reported

1. Low. The adapter rejects attestations whose issuedAt is even one second ahead of the block clock, dropping the protocol library's five-minute drift tolerance. On the paid callback path that single attempt is not retried and the IMD is spent. The same attestation is accepted one block later by manual submission.
2. Info. The hook refuses pool initialization until the owner configures it, so a one-transaction factory launch cannot open the hooked pool. The builder already flagged this. I recorded it because every downstream economic flow depends on that fee stream.
3. Info. Trust assumption: an unbounded owner-set reserve target can route all revenue to the operating reserve and then to the operator, bypassing the 30/30/40 split.
4. Info. The no-argument claimReward() targets the current round, so winners of a finalized round get WrongPhase once the next round opens. The indexed overload works and nothing is lost.

Not reached. No entry point was left unanswered. The live IMD pool's hook behaviour on the IMD purchase route and the production router's handling of hook deltas could not be verified offline and remain owner configuration items, as the README states.