HashDit is a coherent Web3 transaction-safety extension: injects a provider shim that proxies (:47-108), intercepts only, and either replays the ORIGINAL args via Reflect.apply or rejects with providerErrors.userRejectedRequest. It never mutates transaction params, never touches private keys, seed phrases, cookies or localStorage, and the only network destination in the entire reviewed code is the first-party [domain withheld] (HMAC-signed,:530-560, duplicated in:820-878). Two caveats keep this at rather than: (1) is declared in, injected into every page at document_start by:22, listed in the CWS-signed, and is roughly 2.75 MB of the 2.91 MB the ML measured, yet it was absent from the analysis bundle and could not be reviewed, so 95% of the shipped JS is uninspected; (2) a step-4 manifest mismatch exists (bundled:46 declares only "storage" while the live listing declares "tabs"+"storage", and:440 already calls tabs.update / tabs.sendMessage). I am deliberately demoting that to the rationale rather than logging it: the direction is that the reviewed build declares FEWER privileges than the listing, which is version skew on a 160-day-old 0.0.13 snapshot, not a stealth privilege gain, and the `other` category is barred under this verdict. Hygiene, not findings: the API appid and shared secret are hardcoded in cleartext (:547,:857), and every http(s) frame's FULL location.href is uploaded with a persistent storage.local UUID ( URL_CHECK path ->:432-448; UUID minted at:665-675), alongside wallet address, chainId and raw payloads on intercepted calls. Those categories (web browsing history, wallet data) are not covered by the CWS 'none declared' disclosure, which is an under-disclosure problem, but the destination is unambiguously first-party ([domain withheld] matches and the [domain withheld] privacy policy) and the URL check IS the advertised phishing-detection feature, so per the disclosure cross-check it is not a data_exfiltration finding. Corroborating a benign call: the risk engine scored Low, chrome-stats is with a trusted-publisher flag, and the manifest requests only `storage`. Contradicting it: the 10% malware rate in the permhash cluster and the malext-feed 'malware' label, whose own basis text admits it is an unreviewed bulk feed import with no affected version, so it carries little evidentiary weight against the code read here.