Release Tip Jar (ERC-20 symbol TIPS, displayed as $TIPS) on Sepolia chainId 11155111 as an evm_project with one application contract, followed by a publicly hosted, extremely simple website. This is a test launch. Token contract name TipJarToken, name exactly Tip Jar, symbol exactly TIPS, 18 decimals, no constructor arguments, fixed 1 billion tokens (10^27 minor units) minted to msg.sender; no mint/admin/upgrade/tax/blacklist backdoor. Application contract TipVault, constructor (address token) with manifest reference $token, adminless, no owner. stake(uint256) and unstake(uint256) move the launch token in and out (SafeERC20). Anyone can send ETH via tip() or plain receive; each tip is split pro-rata among current stakers using a reward-per-share accumulator scaled by 1e18; tips arriving while nothing is staked are held and credited to the next staker-bearing distribution rather than lost (document the exact choice). claim() pays the caller's accrued ETH with checks-effects-interactions and a reentrancy guard. Views: totalStaked, stakedOf, claimable(address). Events for stake, unstake, tip, claim. No fees, no emissions, no admin, no pause. Constructors take only static words; no post-deploy initialization, no constructor ETH. Solidity with OpenZeppelin, reproducible offline Foundry build with bytecode_hash=none. Tests: pro-rata splits across multiple stakers, a late staker earns nothing from earlier tips, unstake then claim, claim reentrancy attempt fails, zero-stake tip handling, rounding dust bounded, fuzz stake/unstake/tip sequences preserve ETH solvency (contract balance >= sum of claimables). Independent review covers token, application contract, tests and generated launch.json.

Pinned current Sepolia policy v5: 2% rewards this launch's accepted contributors by existing scoring, 8% equally among unique wallets with accepted work in preceding 12 hours, overlap combined into one leaf, 80% LP, 10% treasury, one-hour unlock. The control plane freezes allocations; workers must never invent recipients or choose economic policy. Native ETH pool, fee 3000, tickSpacing 60, policy opening fully diluted valuation 20 ETH for all 1 billion tokens. Policy derives effective price; manifest initialPrice is retained for provenance but overridden by policy. Token-only liquidity, no ETH funding. No mainnet transactions.

After verified deployment, build a minimal React/Vite/TypeScript page with RainbowKit wallet connection, wagmi/viem, and no custom styling (plain browser defaults; RainbowKit's required stock CSS is fine). Offer switch to Sepolia; disable writes when disconnected or on the wrong chain. Show pending, confirmed, rejected-signature and reverted states for every transaction and re-read state after each receipt. Token spends use an exact-amount approve step before the action when allowance is short. Title Tip Jar, display $TIPS and Sepolia. One plain sentence: stake TIPS, anyone can tip ETH, stakers share tips. Show total staked, wallet TIPS balance, wallet stake and claimable ETH. Forms/buttons: stake (approve then stake), unstake, claim, tip ETH amount.

Use actual .imd/reads/deployment.json for token/app addresses, chain, source and ABIs, and build output imd-deployment.json as mandated by the workflow; no invented addresses. Injected wallets must work without operator credentials; do not invent a WalletConnect project ID. Keep web source/package/lock/build config in web/, relative-base static export in root dist/, docs in docs/. Export implementation-derived ABIs and verify canonical ABI hashes against the handoff. Build, typecheck and exercise real browser interactions without broadcasting from worker keys; publish actual evidence and state limitations. A failed RPC read is an error state, never zero. User authorizes normal configured GitHub/IPFS hosting and Sepolia deployment through existing services, not worker access to secrets or direct worker broadcasts. Finish only after source publication, verified deployment, website publication and reachability validation. All contract and workflow steps, including manifest, independent review and frontend, require the premium inference tier with approved highest-capability Claude or Codex models. No downgrade to runtime defaults or economy models.