onyums 0.4.0 — arti 0.46, rustls instead of OpenSSL, and the first verified live serve

The embedded Tor client moves two arti minors and off OpenSSL onto rustls, two security advisories close, and the live onion-service path is verified against the real Tor network for the first time.

Since 0.3.1 (2026-06-18) the published crate has lagged the repository by a wide margin — 0.3.1 predates onyums-skin entirely. This release closes that gap.

Security

0.3.1's dependency graph fails cargo deny check advisories today. This release closes:

Advisory	Crate	Fixed in
RUSTSEC-2026-0258	h2 — unbounded empty DATA frames	0.4.19
RUSTSEC-2026-0285	rustls — TLS 1.3 handshake messages accepted across encryption-level boundaries	0.23.45
yanked	chacha20 0.10.1	0.10.2
RUSTSEC-2025-0052	async-std discontinued	left the tree

arti 0.44 → 0.46, and rustls instead of OpenSSL

arti-client and every tor-* crate move to 0.46.0 (Arti 2.6.0). Notably, 2.5.1 fixed arti reading XON data rates as bits/s rather than bytes/s — sending 8× less data than allowed, squarely on the onion-service path.

The TLS runtime under arti is now rustls rather than native-tls, which removes openssl-sys from the dependency graph: 50 crates and 611 lines leave Cargo.lock, and the binary carries one TLS implementation instead of two.

This is not an FFI-free TLS stack, and the roadmap does not claim it is. rustls' ring provider vendors C and assembly from BoringSSL, so building onyums still needs a C toolchain. The swap replaced one C crypto library with another; what it bought is OpenSSL's removal, an advisory retired, and a single TLS stack. The remaining C dependencies are enumerated and wrapper-pinned in deny.toml.

onyums now installs rustls' crypto provider as the process default before building any TLS configuration, and defers to one your application installed first.

Live verification — a first for this project

Onion service name: xfywsfgmwjvrs3pivzamhjyfxwh6pzhzcgkknpzazhqhqhnygam7u6qd.onion
live tier: fetch attempt 1 served the app body
test result: ok. 1 passed; 0 failed — 257.16s

arti bootstrapped a full directory over rustls, the service published its descriptor, a second Tor client dialled the .onion through a real rendezvous circuit and received the application's body, and shutdown returned cleanly.

Scope of that claim: one host, one run, by hand. CI does not run this tier, so the README's feature-status row stays amber — "verified once", not "continuously verified".

The --ignored live tier itself was also repaired: it could previously hang indefinitely past its own timeouts (observed at 70 minutes against a ~21-minute budget). It now runs on an explicit runtime with bounded teardown under a wall-clock watchdog that aborts and names the stuck phase. On its first real outing it failed loudly in 360s on a cold directory cache instead of hanging — the redesign working as intended.

Fixes

• A peer-provokable memory leak in the accept loop. A rendezvous circuit whose accept() failed was never passed to CircuitPolicy::forget, leaking one per-circuit accounting entry per failed handshake, for the service's lifetime. A peer provokes one by abandoning the handshake at zero cost. (The existing test had pinned the bug by asserting the leak was expected.)
• Two public metrics helpers that misreported under load. circuits_failed_transport() counted host-global capacity refusals as arti transport failures — telling an operator "your Tor transport is failing" when the truth was "you are full". total_streams() omitted capacity-refused streams, so a service serving 1 and refusing 99 computed a 100% served fraction.
• A rustls panic reachable from two public constructors. ProvidedCert::from_pem and the default self-signed acceptor built a ServerConfig before any crypto provider was installed — a panic, not an error, for any downstream app whose graph enables a second rustls provider.

Other

• cargo-deny now denies advisory-not-detected, so an ignore entry whose crate has left the tree fails CI instead of lingering as a decoration — the advisories-side twin of the unused-wrapper guard.
• New troubleshooting entry for arti's fs-mistrust rejecting a world-writable ancestor of the storage directory, which names neither the path nor the remedy on its own.

Breaking changes

• MSRV is 1.91 (from 1.85 at 0.3.1), set by the arti stack.
• OnionTorClient is now TorClient<TokioRustlsRuntime>, was TorClient<TokioNativeTlsRuntime>. Code written against the OnionTorClient alias is unaffected — that is what the alias exists for. Code that spelled the native-tls runtime out will not compile.

Not in this release

onyums-skin is still unpublished on crates.io, so this tag is a GitHub release only; onyums cannot be published to the registry until skin goes first. The full picture of what is verified, what is only offline-tested, and what is blocked upstream lives in README.md and ROADMAP.md.