SolGuardian: multi-agent exploit hunter

Problem. Smart-contract audits are the bottleneck between shipping and being solvent. A human review of a mid-size vault takes days, costs five figures, and still misses things because the same exploit classes (reentrancy, missing access control, unvalidated Solana accounts, manipulable price reads, replayable signatures) must be re-derived by eye in every codebase, across three ecosystems. Attackers automate; new contracts are probed in minutes. Small teams cannot afford audits, so they ship un-audited or drown in linter noise. Solution. SolGuardian is an exploit hunter you point at source. Give it Solidity, Anchor, or a directory; specialised agents cover EVM/HyperEVM and Solana in parallel and return a ranked report. Every finding ships with severity and confidence, plain-English explanation, file:line, educational exploit sketch, a Foundry/Anchor PoC stub, a patch sketch, and the checklist that justified the call. Checklists live in skills/ (reentrancy, access-control, oracle-price, Solana accounts, CPI, HyperEVM notes, severity). Adding a detector means writing its skill first. Design constraint: deterministic code, not a model guessing. Detectors are AST-lite parsers matching only comment- and string-masked source. The report reproduces with zero tokens, runs offline, and works in CI with --fail-on high. Proof. 17/17 seeded issues caught across Solidity, HyperEVM-style Solidity and Anchor; 49 findings with PoC stubs; clean twin contracts report zero findings, enforced by tests. Out of scope: symbolic execution, SMT, live mainnet scanning, credential-dependent features.