[grumbach]

Wire types for clients that read the same file sharing its records over WebRTC, with nodes introducing them and staying the last seed (ADR-0026): the announce, answer and leave messages of a file's swarm room, their shared limits, and the SDP rule both a relaying node and an applying client use so a stranger cannot point readers' connectivity checks anywhere. A second commit makes ChunkMessage::decode refuse bytes after a message. The branch is rebased onto main as three commits, the swarm messages following its get-or-closer request (ant-node ADR-0020). A fourth commit, from review, classes addresses as the IANA special-purpose registries mark them globally reachable (some, such as 2001:4:112::/48 and NAT64's 64:ff9b::/96, had counted as not public, so a node's branch for members on its own network could relay them), drops every candidate in NAT64's well-known and local-use prefixes (64:ff9b::/96, 64:ff9b:1::/48), and adds is_local_ip, the allowlist a node and a client now hold local candidates to, so an address neither public nor local is kept on no network.

One change across six repositories. Merge order:

1. ant-protocol feat(swarm): add swarm rendezvous messages for clients sharing records #50, saorsa-transport feat(webrtc): answer plain STUN Binding requests on the WebRTC Direct socket saorsa-transport#231 and saorsa-core feat(network): send a message on one named channel only saorsa-core#183, independent of each other
2. ant-node feat(swarm): host a swarm rendezvous on every node ant-node#262
3. ant-client feat(swarm): share records with other readers of a file, in every client ant-client#244
4. Try https://github.com/WithAutonomi/try/pull/21, once a browser SDK release carries the swarm

This PR depends on: none.

Linear issue

Closes V2-1892

Risk tier

• T0 — docs / tooling / CI / pure UX-output. Repo CI only.
• T1 — client-only, no network-facing behavior change. CI + prod compat smoke.
• T2 — node/client logic with behavioral surface, no protocol/format/economics change. Dev testnet + ADR.
• T3 — protocol / storage format / payments / routing. T2 evidence + adversarial testing.

Proposed; a human confirms at review.

Compatibility

• Wire: six ChunkMessageBody variants appended (tags 16 to 21, after main's get-or-closer pair); released nodes (v0.23.1, and v0.24.0 of the 2026.11.1 release) cannot decode them, log a warning and leave them unanswered (no panic or ERROR line from v0.23.1 in the rollout lab; v0.24.0 handles an undecodable message the same way), and clients pass those nodes over. ChunkMessage::decode refuses bytes after a message, which no sender writes.
• Storage: none
• API: new swarm module, limits, constrain_swarm_sdp and its address classes (canonical_ip, is_global_ip, is_local_ip); ChunkMessageBody is #[non_exhaustive].

Semver impact

• breaking
• feature
• fix

Test evidence

• cargo test --lib: 132 passed at 4bf8dc9, on main 1179782 (the 2026.11.1 release, ant-protocol 6.0.0, which carries ADR-0020's get-or-closer request): three commits with the same code as the reviewed 5c5a1f1, and the address-class fix. cargo build --no-default-features builds. cargo fmt --check, cargo clippy --all-targets --all-features -- -D warnings, cargo clippy -- -D warnings and RUSTDOCFLAGS="-D warnings" cargo doc --no-deps: clean (Rust 1.99).
• New tests pin every new wire tag, the shared limits, decoding that refuses lists past their limit and bytes after a message, and constrain_swarm_sdp: only the standard form of a candidate or fingerprint line stays (other spellings, Unicode whitespace, stray carriage returns, out-of-range components and priorities go), a mostly dropped SDP holds only what it keeps, and the address classes: public where the IANA special-purpose registries mark an address globally reachable (the first and last address of every prefix the fix changed, 2001:4:112::/48, 192.0.0.9 and NAT64's 64:ff9b::/96 among them, and each one's neighbours), with Teredo, which the registries leave undecided, not public and 6to4, also undecided, public; local only for private use, shared, link-local, loopback and unique local addresses; no listed address both; and no candidate in 64:ff9b::/96 or 64:ff9b:1::/48 kept whatever a consumer allows. The SDP and address tests fail with each fix reverted.
• Adversarial: ant-core parses SDPs this filter let through with webrtc-rs's own SDP and ICE candidate parsers (fingerprints, candidate addresses, candidate priorities).
• Consumers at their PRs' heads: ant-node 48e2761f 1311 lib tests and its WebRTC Direct devnet; ant-core b6f4a8c 989 lib tests, e2e_swarm and 261 WASM tests, all passing.
• CI at 4bf8dc9: every build, test and lint job passes; Security Audit fails. It fails only on RUSTSEC-2026-0387 (crypto-bigint 0.5.5, through elliptic-curve 0.13), which cargo audit reports on this repository's main lockfile too, with the same dependents; its fixes wait on alloy and webrtc-rs releases. linear-link and pr-template, which check this description, pass.

New dependency

none

ADR

https://github.com/grumbach/ant-node/blob/feat/peer-swarm/docs/adr/ADR-0026-clients-share-records-nodes-introduce-them.md (new in the ant-node PR, status Proposed)

Mitigation / rollback

Revert: nothing sends the new messages until the node and client PRs land.

[MaidSafe-Hermes]

REQUEST_CHANGES — 1 P2 source finding

src/swarm.rs:690–695 misclassifies globally reachable IPv6 destinations as non-global. It excludes all of 2001::/23 and does not include the NAT64 well-known prefix 64:ff9b::/96 in the global-address range. The production consumer in ant-node#262 uses this predicate at src/swarm.rs:543–547: when the observed host is local, the local branch retains candidates classified as non-global. An announce can therefore retain 2001:4:112::1 (IANA marks 2001:4:112::/48 globally reachable) or a NAT64 address such as 64:ff9b::808:808, which represents public IPv4 8.8.8.8 under RFC 6052. A receiving client on a NAT64 network can then send ICE checks to that public destination.

Please classify using an explicit, current special-purpose policy, including globally reachable exceptions and translation prefixes, or use a strict allowlist for local candidates. Add regression cases for these addresses and the local-host filtering path.

Panel: 2/2 seats reviewed. sol-1 raised the address-classification issue; glm-1 did not. I confirmed the production call path and checked the IANA IPv6 Special-Purpose Address Registry and RFC 6052. CI checks were read, not rerun locally.

[MaidSafe-Hermes]

REQUEST_CHANGES — reviewed commit 4bf8dc937bd3cb0049ae45fb01491b1c99d5bc0f

P2: shared CGNAT space is treated as proof of a local network. In src/swarm.rs:774–781, is_local_ip returns true for 100.64.0.0/10. The node and client use this result to authorize local SDP candidates. RFC 6598 §1 defines this as shared provider-CGN space between CGN and CPE, not a subscriber LAN; IANA also marks it not globally reachable. A peer's 100.64/10 source therefore does not prove that another member shares its LAN.

Execution path: when a node sees a member at 100.64.0.1, ant-node/src/swarm.rs:545–549 classifies that member as local and retains a 192.168.x.x or .local SDP candidate. A client reached through another shared address then treats reached as local at ant-core/src/swarm/room.rs:1345–1350 and applies that candidate; ICE can probe the receiving client's private network. The supplied tests cover documentation, benchmarking and NAT64 origins, but not shared-CGNAT origins.

Minimum fix: do not let 100.64.0.0/10 alone satisfy the local-network predicate used for candidate authorization. Add tests showing that shared-CGNAT source/reached addresses do not unlock private, ULA, link-local or mDNS candidates; if support is required, use a stronger locality signal than CGNAT membership.

Coverage: 2/2 seats reviewed this exact head's address-classification delta (sol-1 gpt-6.1-sol xhigh; glm-1 z-ai/glm-5.3). Both approved the IPv6/NAT64 classification changes. This shared-CGNAT trust-boundary finding was confirmed separately against the exact callers and RFC/IANA definitions.