[Nic-dorman]

Linear issue

Closes V2-1894

Every e2e test node now binds port 0. Later nodes bootstrap from the addresses the bootstrap nodes actually bound (listen_addrs()). This is the same change as ant-client#248, where the race was found and reproduced. The two retries that worked around it are removed.

The race

• Fixed ports: each test node bound 127.0.0.1 on a port the harness picked, a random base in 20000-60000 plus the node's index. That range overlaps the OS ephemeral range (Linux 32768-60999, macOS and Windows 49152-65535).
• The colliding socket: every node and client starts a UPnP gateway search when its transport comes up (saorsa-transport upnp.rs, enabled by default, in local mode too). The search binds 0.0.0.0:0, igd-next's default, and holds that socket for up to 2 s. A CI runner has no gateway, so it uses the full 2 s.
• Why the bind fails: a 127.0.0.1:P bind while another socket holds 0.0.0.0:P fails with "address in use". When a search socket happened to get the port of a node not yet created, that node failed with Failed to create dual-stack network nodes: Failed to create transport.
• Reproduced on dev2, in ant-client's harness: holding 0.0.0.0:<base+3> gave exactly CI's error string. Details are in ant-client#248, "Flake 2".

What goes

• start_node's 4 attempts on the same port (eb049e3). They fell inside one 2 s search window, so they rarely helped. On ant-client#248, a run failed all 4.
• start_network's bring-up on a fresh port range (627f2eb: with_fresh_ports, NETWORK_START_ATTEMPTS). Its doc says it was also for Windows runners whose reserved port ranges move between runs, where a fixed port can be refused outright. Port 0 covers that case too: the OS never assigns a reserved port or one in use.
• With them: TestNetworkConfig::base_port, TEST_PORT_RANGE_MIN/MAX, the port-overflow check in TestNetwork::new, and TestNode::port. Nothing outside the harness read them.
• The test a_network_whose_port_is_taken_moves_to_a_fresh_range went with the retry it covered. There is no fixed port left to hold. It's replaced by nodes_listen_on_distinct_os_assigned_ports, which checks what bootstrap now relies on: every node reports a non-zero port, and no two share one.

Also changed

• src/node.rs unit tests: five built a node on a random port in the same range. Two retried with a fresh port; three didn't retry. They now use port 0, which NodeConfig::port documents as auto-select. One of them, a_finished_store_survives_a_file_store_that_will_not_open, expects build() to fail, so a port collision could have let it pass for the wrong reason.
• CLAUDE.md: the port-isolation table said tests use 20000-60000. It now says OS-assigned.

Not changed

src/devnet.rs, the devnet behind ant dev start, picks fixed ports the same way (DEVNET_PORT_RANGE 20000-60000). It probably has the same race when it starts many nodes. Not in this PR.

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.

Compatibility

• Wire: none
• Storage: none
• API: none. Test code only (tests/e2e/, the #[cfg(test)] module in src/node.rs) and CLAUDE.md.

Semver impact

• breaking
• feature
• fix

Test evidence

On dev2 (Linux, 4 cores), at 0da08c8 (b0df77a plus this change; dc16dda only edits CLAUDE.md):

• cargo test --lib --features test-utils node::tests: 24 passed.
• cargo test --test e2e --features test-utils -- --test-threads=1, as CI runs it: 121 passed, 0 failed, 3 ignored, the same as main. nodes_listen_on_distinct_os_assigned_ports passed.
• cargo +stable clippy --all-targets --all-features -- -D warnings (rustc 1.99.1, as CI): clean. dev2's default rustc 1.95 flags six lints in files this PR doesn't touch (src/replication/pointer.rs, src/storage/chunk_store.rs); 1.99.1 doesn't.
• cargo fmt --all -- --check: clean.

CI (dc16dda, run 42030617783): green apart from the inherited Security Audit (RUSTSEC-2026-0387, V2-1743). The e2e target passed 121, failed 0, ignored 3 on ubuntu (957 s), macOS (905 s) and Windows (1096 s). Unit tests passed on all three. No "Failed to create node" or "Failed to create transport" in any log, and CI Gate passed.

ant-client#248 made the same change to ant-client's harness and is green on ubuntu and macOS: E2E 69/69 and Merkle E2E 8/8. This PR's CI covers Windows as well.

New dependency

none

ADR

n/a

Mitigation / rollback

Revert. It touches only test code and CLAUDE.md.

🤖 Generated with Claude Code

[MaidSafe-Hermes]

Reviewed exact base b0df77a and head dc16dda. The change is confined to test harness/unit-test code and documentation. Port-0 binding is paired with deriving bootstrap peers from each started node’s listen addresses; the added test checks assigned ports are non-zero and distinct. No actionable defect found.\n\nCoverage: 2/2 seats reviewed (sol-1 gpt-6.1-sol xhigh; glm-1 z-ai/glm-5.3); both seats agree (two model families).\n\nCI on this head: unit/E2E suites pass on Ubuntu, macOS and Windows; CI Gate and other checks pass. Security Audit fails for the inherited RUSTSEC-2026-0387 / V2-1743 issue; this is a separate merge gate, not introduced by this test-only PR. Focused local Cargo test did not complete within the review command timeout.