The review is complete. The findings file holds two substantiated defects, both low severity, and the working tree is otherwise untouched apart from the scratch test.

Contracts read in full: PLEA, CabalGate, Seasons, PleaDistributor, PleaLaunch, PleaHook, Letterbox, OracleAttestation, and both interfaces. Everything in src/ was reached. Not exercised: the fork rehearsals in test/fork/ (no RPC in this environment) and the live Sepolia Intake, Stacker, sIMD and TestIMD contracts, which remain trusted dependencies. The non-fork suite passes (387 tests, 0 failures), the manifest has only the three schema keys with the right argument counts, and the runtime sizes match the task's figures (CabalGate 23,889, PleaHook 24,288, Letterbox 21,383).

Findings, both reproduced by test/scratch/Leads.t.sol:

• Low, PleaHook.sol:655. A 32-byte hookData word with bits above 160 set makes abi.decode(hookData, (address)) revert, so the swap fails instead of falling through to the router's msgSender() or to "no trader" after death. Repro: a buy through a router with hookData abi.encode(uint256(uint160(alice)) | 1 << 200) reverts in both states where the rule says it should go through.
• Low, Letterbox.sol:297. The same-share appeal clamp floors, and the gate's 1% minimum floors from a larger number, so an appeal of a site plea placed at exactly the 1% minimum lands one wei under the minimum whenever the original balance is not a multiple of 100 and the shrunk balance is. appealAuto then reverts AmountTooSmall, about 1 in 100 such appeals; the writer can only fall back to a hand appeal paying IMD. Rounding up or raising the clamp to the gate's minimum fixes it.

Nothing critical or high was found. The new trader rule, keeper-tip removal, fee cap at delivery, Ownable2Step, and appeal-share clamp behave as specified in every other path I traced: custody of escrows and tips, exact-output fee swaps, PLEA transfer gating, Seasons accounting with ERC-6909 claims, and the hook's delta settlement in both pool orientations.

Trust assumptions to carry forward: the oracle signer and Intake (the gate binds neither a.requestId to the Intake id nor questionHash to the body, so a misrouted callback from the Intake would be accepted), the Stacker's credit, the PLEA owner's add-only allowlist, the distributor owner's one-time Merkle root, and the immutable hook fee beneficiary.

Fixed both audit findings and the related capped-appeal minimum bug. Added regressions, updated the Letterbox hash, and documented changes in ADAPTATION.md.

forge build passed. Tests: 407 passed, 0 failed, two fork suites skipped because RPC URLs were unset.

Launch gas with 22% headroom: 15,436,103, below the cap. All contract sizes remain within limits.

The review is complete and recorded in .imd-findings.json with 4 findings and 50 coverage rows, one for every listed entry point.

What I found in the permissions area

• Low. Season exclusion follows the mutable owner, the fee beneficiary does not. Seasons excludes plea.owner() at record and close time, while the hook's 0.5% fee recipient and PleaLaunch's owner are immutable copies taken at launch. After a two-step handover to the team's second wallet, the original wallet keeps every owner fee and also earns points, the IMD prize and the 1,250,000 PLEA bonus. A fixture test reproduced the full sequence.
• Low. The gate never pins the signed questionHash. onOracleResult trusts the Intake's intakeId routing and checks only signer, replay, panel and range. A valid attestation signed for plea A's question, delivered under plea B's intake id, approved plea B with score 45. The attacker here is a dishonest or buggy Intake, so the severity stays low.
• Info. hookData writes a stranger's buy record. A buy that names another wallet moves that wallet's buy-weighted holding time toward now, which is a fact-score input. It costs the attacker the IMD and gifts the PLEA, so it is a documented asymmetry, not an extraction.
• Info. Owner allowlist as a trust assumption. Allowed senders can move PLEA off-court for good while everyone else must plead. Recorded as the intended power the brief asks for.

What held after tracing the guards, state changes and failure paths: launcher-only wiring and one-shot initialisation in PLEA and PleaLaunch, the Letterbox-only gate paths and courtTake, the PoolManager-plus-flag callback guards in all four unlock callbacks, sells only through the gate while the Cabal lives, the no-tip rebalance, and the Merkle distributor. The manifest test confirms the committed Letterbox hash and the registration window.

Not reached: the mainnet fork rehearsal (no RPC in this sandbox) and a candidate about contracts under construction posing as plain wallets, which I dropped because such a contract cannot receive the PoolManager's unlock callback. Scratch tests live in test/scratch/AccessReview.t.sol and all three pass on the current code.

Findings file written and validated. The working tree shows no tracked changes, and the scratch directory is removed.

Outcome. I traced the control flow of all 54 listed entry points across CabalGate, Letterbox, PLEA, PleaDistributor, PleaHook, PleaLaunch and Seasons, plus the launch manifest, against the Execution Trace, Periphery and First Principles checklists. I found no exploitable defect. The full non-fork test suite passes on this tree. The Letterbox creation-code hash compiled here matches the value in launch.json, the manifest schema and constructor argument counts are correct, and the launch fits the gas cap.

Reported in .imd-findings.json (both info severity, with concrete reproductions):

• Stale gas figures in the manifest notes. The notes say 11.6M gas and 14.1M with headroom. The repository's own LaunchManifest test measures 12,652,544 and 15,436,103 on this tree. Still under the cap, but the stated margin is roughly double the real one.
• Runtime size budget drift. The brief fixes PleaHook at 24,288 bytes and Letterbox at 21,383. This build gives 24,296 and 21,397, so the hook's EIP-170 margin is 280 bytes, not 288. CabalGate matches exactly.

Leads checked and rejected. The trader rule's "plain wallet" test could in principle match a contract still in its constructor, but the PoolManager cannot call back into an address without code, so the path is unreachable and I did not report it. The run-4 changes (trader resolution, keeper tip removal, the capped delivery fee, the appeal fee and share-band clamp, Ownable2Step) all behave as specified along every path I traced, including the boundary between the 7-minute sale window and settle, and the consistency of the three staleness checks between the gate and the Letterbox.

Not covered. The fork tests (test/fork) need MAINNET_RPC_URL and were not run. The Sepolia addresses in launch.json (Intake, oracle fee asset, TestIMD, signer, PoolManager, Stacker) were not verified against the live chain.

The review is complete. The findings file is written and no tracked file was changed.

Outcome. One substantiated defect, severity low, in the Letterbox's appeal path. Everything else in the math, boundary and numerical-gap area held under the checks I traced.

Finding 1 (low), src/Letterbox.sol line 306. The run-4 change (d) clamps an APPEAL letter to the original plea's share of the balance. That clamp always produces an amount smaller than the original by the fee sale's share, and the next line then applies the same 0.70 IMD floor the original had to clear. A 20% letter worth between 0.70 and about 0.75 IMD at the snapshot is accepted as a plea but its appeal is refused with TooSmallForTheFee, by both the APPEAL letter and appealAuto. The gate accepts the identical amount through CabalGate.appeal, so no funds are at risk. The scratch test in test/scratch/AppealFloor.t.sol fails on this tree with that selector and is attached as the proof. Reproduction numbers are in the file.

What held. The hook's two-stage fee (specified-side fee reconciled against the fill, parts never exceeding the fee), the cap decay remainder arithmetic, tick alignment for negative ticks, the seed price in both orientations, the Seasons pot accounting invariant through close and flush, the prize shares summing exactly, the gate's 1%/50% share bands on the lesser balance, and the Letterbox's custody and capped delivery fee. The six existing suites nearest my area pass, 83 tests.

Coverage. All 48 listed entry points have a row, plus two invariant rows. Three rows point at finding 1. None are marked unreached. Entry points outside my area are marked holds with the reason that nothing numerical touches them.

The review is complete and the findings file is in place. Nothing else is pending.

Outcome. The economics of PLEA run 4 hold up under the Economic Security, Invariant and Flow Gap checklists. I found one low-severity gap, verified it with a Foundry test, and recorded coverage for all 52 listed entry points plus four invariants and the manifest. The findings file is at .imd-findings.json; the working tree has no other changes (the scratch test lives under test/scratch/, which is discarded).

Finding (low, verified). The owner exclusion from points, prizes and the PLEA bonus is checked only against whoever owns PLEA at the moment of record or closeSeason. The owner wallet can plead, get approved, rotate ownership to a second team wallet with the two-step transfer, and then be recorded and paid the IMD share plus the 1,250,000 PLEA bonus share. The scratch test reproduces exactly that payout on this tree. This is reachable only by the privileged owner, so it is reported as an unenforced trust assumption against the brief's "leaves the team for good" promise, not as a third-party bypass.

What held. I traced the Letterbox escrow and tip custody in both fee modes, the exact-output fee swap and its cap, the appeal clamps added in run 4, the deliver fee cap, Seasons pot accounting and rollover, the hook's fee allocation, claims, wall reserve and the new trader rule, and PLEA's transfer gate. The removed keeper tip is gone from the source. The manifest has the right shape and argument counts, and the manifest test suite passes on this tree, including the committed Letterbox hash and the gas cap.

Not reported. Two leads were dropped as accepted design: attestations are not bound to the plea's intake id in the gate (the repo's own replay test states the Intake owns that binding), and hookData lets a buyer name any non-excluded address as trader, which only costs the buyer.

Limits. The full build took most of the budget, so I did not re-run the whole suite or the fork test. Static-analysis leads in my area (strict equalities, reentrancy flags) were checked by reading and did not reproduce as defects.

Added failure-path tests and stateful custody invariants for trader resolution, gas limits, fee surplus, vault discovery, and ownership transitions. Changes are limited to test/.

Validation: forge build passed; forge test reported 441 passed, 0 failed, 2 skipped. Fork tests lacked RPC URLs.

No new defects reproduced.

Updated launch.json with run-4 notes and verified gas figures. Constructor arguments and Letterbox hash match the accepted build.

forge build passed. forge test: 407 passed, 0 failed; 2 fork checks skipped without RPC URLs. Only launch.json changed.

The review is complete and .imd-findings.json holds six findings plus a 48-row coverage record with no unreached rows. No repository file was changed; only the findings file and three scratch tests under test/scratch/ were written.

What I kept, all reproduced on this tree

• Low, Letterbox line 306. An APPEAL letter of a floor-sized 20% letter is refused. The run-4 same-share clamp always makes the appeal smaller than the original by the fee sale's share, then the same 0.70 IMD floor is applied again. A writer holding 3.5 to 3.75 IMD worth of PLEA can plead by letter but not appeal by letter, while the gate route accepts the same amount. The specialist's proof fails here with TooSmallForTheFee, and it is attached as the proof.
• Low, Seasons line 154. The owner exclusion follows the PLEA owner of the moment, while the hook's fee beneficiary is immutable. After a two-step handover inside the team, the former owner still earns the 0.5% fee and now records points and is paid the IMD prize and the 1,250,000 PLEA bonus. I merged the permissions and economics reports into one finding and attached a scratch proof that fails on this code.
• Info, three trust notes. The gate never binds a signed attestation to its plea (the Intake's routing decides), hookData lets a stranger rewrite another wallet's buy record (fact score 42 to 36 in the test), and the owner's add-only allow exemption. All three reproduce in scratch tests but rest on the agreed design or trusted IdentityMD contracts.
• Info, sizes. PleaHook is 24,296 bytes and Letterbox 21,397, 8 and 14 bytes over the brief's budget. Both are under EIP-170, with 280 bytes of hook margin.

What I dropped. The gas-figure note on launch.json does not reproduce: the notes already state 12,652,544 and 15,436,103, and the manifest suite logs exactly those numbers.

Other checks that passed. The Letterbox creation-code hash in launch.json matches this build, the manifest suite passes on an empty Sepolia chain, and the manifest's shape, argument counts and constructor dependency order are valid. The full forge build --sizes was killed at the 20-minute limit on this 4-core machine, so sizes were read from the sparse-compiled artifacts instead; this is the one verification that was not done the way the brief names it.