shoes/DEPENDENCIES.md

Dependency Maintenance

Refresh registry dependencies and maintained build tools without changing peer fixtures. Change supported Rust/platform floors only when explicitly requested. Use targeted Cargo updates and inspect source changes as well as versions. Existing wildcard requirements are intentional; narrower requirements must be reconsidered when upstream advances. Regenerate dependency notices after lockfile changes.

Protected Sources

The workspace patches in Cargo.toml are required by checkout/release builds:

• cfal/rustls: REALITY/TLS integration; retain the matching exact rustls requirements in both workspace manifests.
• cfal/BLAKE3: binary-context derivation for VLESS Encryption.
• cfal/h3: empty-DATA handling and idle-stream cancellation, used by both Shoes and Hickory through h3/h3-quinn.
• cfal/hickory-dns: connection-close recovery and healthy-upstream failover.
• cfal/smoltcp: TUN/WireGuard TCP retransmission ACK sequencing.

Preserve each repository and full revision unless explicitly updating that fork. Check every resolved consumer, especially Quinn, tokio-rustls, hyper-rustls, rustls-platform-verifier, Hickory and Shoes; absence of unused-patch warnings is not sufficient. Keep publish = false while workspace patches remain.

smoltcp

The October 2026 refresh rebases the two TCP recovery commits from published cfal/smoltcp main (a80fa50dce926117669ef061f55f4d6bdad0904a) onto upstream 24bda3b4f982625d8453d4360172557d24921c76. Their previous upstream base was 72a0d7892d696620541b1147c7cac471d421e554, not the older release-based fix/tcp-ack-sequence branch. The new series lives on shoes/tcp-recovery-20261028-GsfiI8, pinned at 042e568bfa39d0f111a23a4718ec257e474fd1e7; existing branches and old pins remain intact.

The package version and latest registry release are still 0.14.0. The new upstream change is unreleased smoltcp #1223: defensively clamp the remaining advertised receive window after FIN consumes sequence space at a full window. Earlier post-release RTO bounds, zero-window timer recovery, failed-fast-retransmission retry and no-std CUBIC fixes were already in our pin and remain included. This refresh changes no public API, features, dependencies, Rust minimum or platform support.

Preserve the complete recovery patch, not only its persistent send frontier:

• Pure ACKs use the non-rewound frontier; a zero peer window with queued stream data uses SND.UNA, while FIN-only flight retains the frontier.
• RSTs retain the baseline retransmission cursor. Fast retransmit cannot extend the frontier with unsent data. Keepalive probes use SND.UNA minus one.
• Probes and empty emissions do not advance the retransmission cursor; failed emission does not advance the send frontier. Socket reuse resets it.
• Relaxed receive acceptance applies only to empty None/PSH segments, bounded by scaling slack, physical receive capacity and the one-byte lost-FIN allowance.

The exact lost-FIN allowance, FIN-only ACK frontier and RST cursor are retained fork policies, not universal RFC or OS behavior. Keepalive sequence placement does not establish timer reachability with outstanding data.

Shoes retains explicit CUBIC, IPv4/IPv6, medium-IP and 32-segment assembly. Linux TUN transmit offload remains IPv4-only: the interface still does not forward IPv6 segmentation metadata. TUN supplies partial TCP checksums when offload is enabled; WireGuard's queue device uses software checksums and MTU-sized packets without segmentation capabilities. Do not enable new offload paths without checking Device, tokens, PacketMeta, ownership and poll wakeups.

For future refreshes, discover and compare published fork branches before rebasing, inspect the complete upstream delta and range-diff, and pin the tested full SHA in both manifest and lockfile. Run the fork's ci.sh matrix, socket::tcp::test:: with Reno/CUBIC, and library tests with the exact Shoes feature set. Use separate target directories for baseline comparisons. The oversized-frame offload fixture selects no congestion control; a separate CUBIC regression checks its real flight limit. This avoids mistaking the old fixture's 65 KB expectation under a 2,048-byte CUBIC window for a runtime bug. Require deterministic loss/window/FIN/reset regressions as well as Shoes' rootless TUN, shared IP-stack, WireGuard, kernel and macOS rootless-TUN checks. Neither those tests nor the original Shoes #197 establish the historical macOS stall's cause or comprehensive Linux/BSD/XNU interoperability.

nix

The nix-rust/nix patch is upstream, not a maintained fork. Published 0.31.3 still lacks #2821, despite sharing the pinned commit's package version. With libc 0.2.191 and Shoes' iOS features (fs,ioctl,net,socket,time,uio,user), registry 0.31.3 fails to compile on aarch64-apple-ios; the pinned commit succeeds. Retain it until a published release passes that regression and the complete mobile builds.

October 2026 Refresh

Registry updates include notify 8.2.1, Web PKI roots/root-certs 1.0.11, ICU 2.4, ctutils 0.4.6, der 0.8.3, cc 1.10.0, openssl-probe 0.2.2, smallvec 1.16.4, tinystr 0.8.5, uuid 1.29.0, yoke 0.8.4, zerovec 0.11.9/derive 0.11.7 and rustix 1.1.6 through the test-only tempfile dependency. ctutils fixes RUSTSEC-2026-0385; the root set removes two more eMudhra anchors. ICU updates Unicode/IDNA data, and der tightens malformed DER handling. Notify changes the macOS FSEvents bindings; stream-wide restart errors must not be swallowed as optional per-path failures. See the changelog for user effects.

parking_lot stays on 0.12.5 (core 0.9.12). Version 0.13/core 0.10 probes futex_time64 before the legacy syscall for 32-bit waits and wakes, falling back only on ENOSYS. Android API 26-29 app seccomp instead traps syscall 422 with SIGSYS; API 30 and later explicitly allow it. There is no published fixed successor. A global 0.12 constraint preserves the API 21 floor without adding platform-specific lock types or rewriting synchronization. Reconsider when an Android-safe stable release exists, including 0.13's removed eventual fairness and the public DynamicFileLogWriter::new mutex-type compatibility. See the upstream probe and Android 8 ARM policy.

The other non-excluded direct dependencies were already at their latest non-yanked stable releases, including AWS-LC, Quinn, GotaTun, tun, rcgen, instant-acme, Tokio and ML-KEM. Prereleases (including notify 9) and unreleased git commits are not stable upgrades.

Newer incompatible transitive families remain constrained by their parents: GotaTun retains its older RustCrypto/rand/x25519/thiserror families; Hickory retains prefix-trie 0.8 and system-configuration 0.7; Quinn retains quinn-udp 0.5; DashMap retains hashbrown 0.14; prost-derive caps itertools at 0.14; ASN.1 parsers retain nom 7 and num-bigint 0.4. Platform parents likewise constrain older Windows/JNI/WASI families. Do not force these through patches, extra direct dependencies, or fork-manifest edits. The deprecated serde_yaml dependency still pulls unmaintained unsafe-libyaml; replacing the configuration parser is a separate compatibility change, not an available version bump.

Build tools use cargo-about 0.9.3, cargo-ndk 4.1.2, cross 0.2.5, Node 26 and the Rust 1.99.1 Docker builder. The Rust minimum is now 1.99.1, checked separately in CI, and the Docker runtime remains Bookworm. Go 1.27.3, cbindgen 0.29.4 and cargo-llvm-cov 0.9.1 were already current. Keep peer hashes, frozen Shoes baselines, fixture-generation provenance and historical benchmark evidence unchanged when refreshing tooling.

Validation

Use the test guide and current workflows for exact commands. In addition to formatting, strict clippy, unit/process and license checks, run the pinned REALITY/Vision, VLESS Encryption, XHTTP/H3, Shadowsocks, DNS and WireGuard suites. Preserve binary-context vectors, H3 empty-DATA/idle-stream cancellation, TLS authentication and TUN/WireGuard loss-recovery regressions. Require nonzero test counts and verify that requested peers actually execute.

Hosted macOS watcher/rootless-TUN tests, desktop builds, Android arm64/armv7 and iOS device/simulator builds remain required. Cross-compilation is not device runtime coverage; Windows and mobile-device execution are not covered by these workflows. Do not claim those platforms passed runtime tests.