Content-Blind by Construction: Building, Auditing, and Legally Situating a Zero-Budget End-to-End Encrypted Messenger

End-to-end encryption is a solved engineering problem; operating an end-to-end encrypted service is not. We present NarChat, a self-hosted encrypted messenger built from scratch on a zero budget — one human directing, AI agents writing and adversarially auditing the code, consumer hardware in a private home — and we use it to examine why perfect communication privacy remains rare when its mathematics are free. The system provides content confidentiality by construction: a hybrid X25519/XSalsa20-Poly1305 fan-out scheme, a challenge–signature login designed so that the password never leaves the device — the pilot's incomplete migration to it is reported as honestly as the design — and a server reduced to a courier of opaque blobs — simple enough that its strongest privacy claim is verified by a test that reads the server's own message storage from disk, searches for a sentinel plaintext, and fails if it finds one. We report with equal prominence what the system does not provide — forward secrecy, metadata privacy, encrypted call signaling — and the claim's scope: confidentiality holds against a passive or storage-reading server, while an actively malicious server could attack key distribution or call signaling and be caught only by out-of-band verification. Independent adversarial audits during development caught real vulnerabilities — a path traversal, a pre-authentication denial of service, a CSRF quota drain, an XSS scheme bypass — each fix locked by a regression test demonstrated to fail on the unpatched code. A pilot deployment (8 accounts, 41 messages, 14 days of uninterrupted operation at measurement) confirms that the engineering works — and that adoption does not follow automatically. We then map the system against Turkish law, as a case study of one jurisdiction, and find the pressure point is not where engineering intuition expects: the statutes' operative demand is traffic data, not message content. The codebase accordingly forks: the full-E2E reference is frozen and archived, and a compliance-instrumented variant has been built against it (not deployed — going live remains a human decision); the complete difference between the two, measured line-by-line, is 91 lines added and 2 removed in one file, with the end-to-end core bit-for-bit unchanged — the fingerprint a jurisdiction leaves on an architecture. We conclude that jurisdiction is a design parameter — as real as bandwidth. This record contains the paper in English (canonical) and Turkish (claim-synchronous translation). The system it documents is archived separately as software, with its full git history, at 10.5281/zenodo.22017587, and is developed at github.com/NarchTech/narchat. Every technical claim in the paper binds to an artifact in that archive — a file and a line, a named test, an audit-log entry — never to a running service. Conventions: claims that could not be verified against a primary source are marked in place rather than smoothed over (rendered as [!] in these PDFs). Legal material is reported, not adjudicated: the paper states what the statutes and case law say and marks what remains untested. It is not legal advice. Status: preprint. This version has passed claim-level verification against the frozen archive and primary legal texts, adversarial multi-model review of its highest-risk sections, and three independent fresh-eye reviews. Further polishing rounds are planned; Zenodo versioning means the concept DOI always resolves to the latest version.