GREEN. rustls 0.23 selects its CryptoProvider from its own 'ring' /
'aws-lc-rs' crate features. Arti reaches rustls through tor-rtcompat with
default-features = false and nothing in this workspace turned a provider
feature on, so rustls was compiled with zero providers and the first TLS
config built from the process default panicked:
Could not automatically determine the process-level CryptoProvider from
Rustls crate features. ...
That is the 'Failed to start' screen on the unlock path. The 'ring' crate
being present in the graph (via snow, for Noise) never had anything to do
with it.
'ring', not 'aws-lc-rs': aws-lc-rs wants CMake and a C toolchain for the
NDK and fights the cross-compile.
Belt and braces: open_wire() now calls install_crypto_provider() before
Node::start_with_passphrase, a OnceLock-guarded
ring::default_provider().install_default() with the already-installed Err
ignored. That keeps the cdylib correct regardless of how a consumer's
feature graph resolves rustls.
aws-lc-rs is absent from the lock, so there is no two-provider conflict to
report.
Evidence:
cargo test --release --test provider_resolution
before: panicked at rustls-0.23.44/src/crypto/mod.rs:249 (device message)
after: 1 passed
cargo tree -f '{p} [{f}]' -p rustls
rustls v0.23.44 [log,logging,ring,std,tls12]
RED. The APK builds, installs and opens, but the first rustls config built
from the process default panics one call deep inside Arti's rustls backend:
Could not automatically determine the process-level CryptoProvider from
Rustls crate features.
Call CryptoProvider::install_default() before this point ...
rustls 0.23 picks its provider from its own 'ring' / 'aws-lc-rs' crate
features, not from which crypto crates happen to be linked. Arti reaches
rustls via tor-rtcompat with default-features = false, so neither provider
feature is on and rustls compiles with zero providers. 'ring' does appear in
cargo tree, pulled in by snow for Noise -- unrelated and irrelevant.
This commit adds rustls as an explicit dependency with no provider feature
(no behaviour change: it is already in the graph that way) so the test can
name the type, plus the regression test. On the host it reproduces the
device's panic verbatim.
The SDK lockfile is resynced with the root Cargo.toml on main (onionwire
0.2.1, rpassword, md-5, rtoolbox) in the same commit so the test build is
against a consistent lock.
