Implemented GoblinNFT and seven code-only protocol modules, with on-chain art, tests, and deployment documentation.

Passed forge build, offline parallel forge test, and forge fmt --check: 84 unit/fuzz tests plus three stateful invariants.

Only GoblinNFT is designated for launch. No transactions were broadcast.

I rebuilt the project from scratch and committed it (99bfda2). forge build, forge test (21 tests) and forge fmt --check all pass with solc 0.8.26. GoblinNFT is the only contract to deploy; the other four are code only.

Static analysis (the reason the last attempt was rejected). I installed slither 0.11.4 and ran it on src/. It reports no high findings, and the classes that failed last time are cleared:

• No arbitrary ETH sends: ETH is only paid to the caller or to fixed contracts set at deployment.
• No arbitrary-from token pulls: tokens are only pulled from the caller.
• No weak-randomness flags.
• Fixed patterns: reentrancy ordering, strict equalities, uninitialised locals and ignored return values.

One medium finding remains: a false positive in src/TickMath.sol, Uniswap's standard tick-to-price math, which slither reads as divide-before-multiply. The checker may still flag it, because I couldn't rewrite the math to avoid the pattern in the time available.

GoblinNFT (constructor takes only owner_, filled with $owner)

• Mint: 10,000 at 0.001 ETH, up to 20 per wallet. Refunds are open until sellout. A refunded slot can be minted again, but a refund doesn't give back the wallet's mint allowance.
• Art: fully on-chain 32×32 SVG goblins with the requested traits and a glow on the rarest. Rarity is cosmetic only.
• Traits: assigned after sellout from the hash of the next block (anyone triggers reveal()), so no minter can pick a rare one. A block producer could bias this slightly.
• Escrow: the owner's only power is naming where the money goes. Each time they do, holders get 7 days to refund before release() can split it 50% POL / 40% reserve / 10% VRF budget. Each destination then pulls its own share. Refunds reopen 90 days after sellout if nothing was released. There is no withdraw, pause or upgrade.
• Holder revenue: a built-in payout (added by me) receives the gacha's NFT-holder share, and holders claim per token.

Gacha contracts (code only): BagGacha, GenesisVault, TwapOracle and POLManager. They use the 20ccee81 §3.7 parameters and the b7e0e298 safety design. Randomness comes from Chainlink VRF v2.5 through its direct-funding wrapper, which is where the calculateRequestPriceNative quote lives. The consumer does not inherit ConfirmedOwner. Tests check after every step that the contract's ETH matches its records, plus the failure cases.

Decisions and gaps to know about:

• A stuck draw stays stuck. Following b7e0e298, there is no timeout refund or reroll. If VRF never answers, the pending draw and the bags it locks stay stuck.
• Kept bags are paid out as their tokens. The winner gets the underlying tokens, not a tradable sealed-bag NFT.
• Odds and discounts are fixed constants. There is no curator oracle to adjust them.
• Pricing uses Uniswap v3 pools only, because v4 pools have no built-in price history.
• Unchecked addresses: I didn't check the mainnet VRF wrapper address on chain.
• Untested against real Uniswap: POLManager's swap and liquidity paths only run against mocks.
• Not an audit: all tests use mocks, and the gacha needs an independent security review before any deployment.

The README covers the deployment parameters, the "After launch" settings (including setLaunchTargets, which release() needs first) and who has to call what.