fix(deps): migrate to stable pkcs5 / pkcs8 / ed25519 and loosen prerelease pins (extends #697) - #702

Context

After pkcs8 0.11.0 stable landed (2026-04-28, see #697), russh's prerelease crypto stack falls over for any downstream consumer that pulls in rsa = "^0.10.0-rc.18". Issue #697 was closed by 2a49916 with a stopgap that pins every prerelease dep with =, releasing as 0.60.2. That fixes russh in isolation but blocks downstream consumers: their newer rsa pulls in pkcs8 0.11.0 stable through the dep graph, which makes 0.60.2's rsa = "=0.10.0-rc.16" and pkcs8 = "=0.11.0-rc.11" unresolvable in the same workspace.

This PR completes the work of #697 by migrating russh's source to the stable pkcs5 / pkcs8 / ed25519 APIs and loosening the prerelease pins so the resolver can climb to versions that compile against the stable formats family.

Change

Source change in russh/src/keys/format/pkcs8.rs:

- pkcs5::pbes2::Parameters::pbkdf2_sha256_aes256cbc(rounds, &salt, iv)
+ pkcs5::pbes2::Parameters::generate_pbkdf2_sha256_aes256cbc(rounds, &salt, iv)

The pbkdf2_sha256_aes256cbc constructor existed in pkcs5 0.8.0-rc.13 and was renamed to generate_pbkdf2_sha256_aes256cbc in 0.8.0 stable (per the pkcs5 CHANGELOG and current docs.rs). Same arguments, same crypto semantics.

Floor bumps in russh/Cargo.toml:

The = prefix is dropped on every line above. Each new floor is the first version of the crate that compiles against the stable formats family. The cascade is forced because rsa 0.10.0-rc.16 source itself fails against pkcs8 0.11.0 stable with the same KeyMalformed enum-variant signature change that broke ed25519-rc.4 in the #697 report, and primefield 0.14.0-rc.7 (pulled in via p256/p384/p521 0.14.0-rc.7) fails against crypto-bigint 0.7.3 with const-generic type-inference errors.

Other prerelease pins (aead, aes-gcm, curve25519-dalek, ml-kem, the internal-russh-num-bigint fork, pkcs1) are left untouched. Each of those crates' latest published version equals the pinned version today, so loosening would be a no-op for resolution.

Verification

On the fork branch with rustc 1.88.0 (the rust-toolchain.toml floor):

cargo build --workspace                                   clean
cargo build --workspace --all-features                    clean
cargo clippy --workspace -- -D warnings                   clean
cargo clippy --workspace --all-features -- -D warnings    clean
cargo fmt --check                                         clean
cargo test --workspace                                    76 lib tests + all
                                                          submodule and doc-tests
                                                          pass
cargo test --workspace --all-features                     77 lib tests + all
                                                          submodule and doc-tests
                                                          pass

cargo minimal-versions check --all-features --no-dev-deps was not run locally (cargo-minimal-versions and cargo-hack are not installed in the development environment). The new floors are chosen so the lowest version satisfying each loosened spec is the first version that compiles against the stable formats family, which is what minimal-versions selects in the absence of an = pin. CI is the canonical check.

Downstream evidence

The fix unblocks at least one downstream consumer: the rustfs SFTP server. The rustfs feature branch carries:

[patch.crates-io]
russh = { git = "https://github.com/simon-escapecode/russh", branch = "fix/697-migrate-stable-crypto-stack" }

against rustfs's russh = "0.60.0" workspace dep. With this patch in place, rustfs resolves cleanly against rsa 0.10.0-rc.18, pkcs5 0.8.0 stable, pkcs8 0.11.0 stable, ed25519 3.0.0 stable, and the rest of the stable formats family. Without the patch, the rustfs build fails on the same KeyMalformed signature mismatch in rsa 0.10.0-rc.16 source that blocks anyone else trying to consume rsa rc.18+ alongside russh 0.60.2.

Risk

The bumps ed25519-dalek 3.0.0-pre.6 -> 3.0.0-pre.7, ecdsa rc.16 -> rc.18, elliptic-curve rc.28 -> rc.32, and p256 / p384 / p521 rc.7 -> rc.9 are minor RustCrypto patch-level moves. The russh test suite covers the affected key handling and ECDSA / EdDSA paths and passes on the fork at every step. The rsa = "=0.10.0-rc.16" -> "0.10.0-rc.18" move is the deliberate fix for downstream compatibility; the = was the root cause of #697's downstream impact.

If cargo minimal-versions check --all-features --no-dev-deps flags any further floor that the in-tree CI catches, the fix shape is the same: bump the floor to the first version that compiles against the stable formats family. Happy to amend.