[Nic-dorman]

Follow-up (c9b13f4): Hermes's non-blocking nit. A comment in e2e_cost_estimate.rs still said 10 nodes. Comment only; optional re-review ac94cd2..c9b13f4.

Round 2 (ac94cd2): test nodes bind port 0, which replaces the node-create retry from round 1. See flake 2 below. Review range dec642c..ac94cd2.

Fixes the two E2E flakes behind V2-1852:

1. "need 7": all 4 fail-then-pass macOS E2E runs in the ticket.
2. "Failed to create transport" when a test node is created. It hit this PR's own CI twice, and twice before (10-23 and 10-29).

The ticket's fifth run, Merkle ubuntu with "Got 14 merkle candidates, need 16", is a different failure. It's now tracked in V2-1893.

Flake 1: "need 7"

All 4 fail-then-pass macOS E2E runs in the last 50 were one test, e2e_cost_estimate::test_estimate_all_stored_partial_sample_is_incomplete:

• Runs: 40015142708, 40349637751, 40683863980, 40844173225.
• Failure: each failed its upload before payment, with InsufficientPeers … witnessed close group initial lookup found 6 peers, need 7.

Cause: the test harness gives every DHT request 5 s. saorsa-core uses a node's connection_timeout for two things: the connection timeout, and the request timeout of every DHT operation (network.rs: request_timeout: config.connection_timeout). MiniTestnet set it to 5 s, against 25 s in production. On a loaded macOS runner, a slow but live peer that answers after 5 s isn't counted, so the lookup comes up one short.

The 5 s came in with the saorsa-client port (5d25103), with no comment. ant-node's harness has the same trap at 2 s, set to keep teardown short (857398c).

Flake 2: "Failed to create transport"

spawn_node panicked with create P2P node: Transport(SetupFailed("Failed to create dual-stack network nodes: Failed to create transport")) before any test logic ran:

Run	Date	Job	Test
39942512923	10-23	Merkle E2E macOS	test_external_merkle_partial_payment_is_partial_upload
40914982712	10-29	E2E macOS	test_attack_zero_amount_payment
41883601747	11-04 (this PR)	E2E ubuntu	test_file_download_bytes_written, the first test in its binary
41897795306	11-04 (this PR)	E2E macOS	test_idempotent_chunk_storage, all 4 retries from round 1

Cause: a UDP port collision with the UPnP gateway search.

• 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 49152-65535).
• 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 SearchOptions) and runs for up to 2 s (DEFAULT_DISCOVERY_TIMEOUT). There is no gateway on a CI runner, so it uses the full 2 s.
• A 127.0.0.1:P bind while another socket holds 0.0.0.0:P fails with "address in use". So when a search socket happened to get the port of a node not yet created, that node failed.
• saorsa-core formats the error with {} (transport_handle.rs:545), which drops the underlying io::Error. That's why the log only says "Failed to create transport".
• Round 1's retry used the same port 4 times over about 1.5 s, inside one search window, so it failed all 4 times on 41897795306.

Reproduced on dev2:

• Base port pinned (local patch only), 0.0.0.0:<base+3> held by a plain UDP socket.
• Result: create P2P node after 4 attempts: Transport error: Setup failed: Failed to create dual-stack network nodes: Failed to create transport, the same string as run 41897795306, in 4.8 s.
• Sampling the test process's sockets with ss showed each node holding only its own 127.0.0.1 socket, and one short-lived 0.0.0.0:<ephemeral> socket appearing with each node spawn.

Fix: nodes bind port 0, so the OS picks a free port. The OS never assigns a port that is in use, whatever holds it: a search socket, another process, or an IPv6 dual-stack socket. Later nodes bootstrap from the addresses the bootstrap nodes actually bound (listen_addrs()). The retry is removed: it only covered this race, and any other failure should show on the first attempt. Nothing else read the fixed ports; no test restarts a node on the same address.

Changes

1. TEST_CORE_CONNECTION_TIMEOUT_SECS = 10, with a comment explaining the double use. ant-node moves to the same name and value in its own PR (V2-1869), and each harness says to keep the two in step.
2. teardown() shuts nodes down concurrently, as ant-node's does, so the DHT leave waits overlap instead of adding up.
3. e2e_cost_estimate.rs uses DEFAULT_NODE_COUNT (14) instead of a hard-coded 10. It was the only e2e file left on 10, and the merkle test in it keeps 20. This helps a little; change 1 is the actual fix.
4. spawn_node binds port 0 (round 2), and start_inner builds the bootstrap list from the nodes' listen_addrs(). This removes TEST_PORT_RANGE_MIN/MAX, the random base port and the 4-attempt node-create retry.

Linear issue

Closes V2-1852

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 harness only (ant-core/tests/).

Semver impact

• breaking
• feature
• fix

Test evidence

All runs below were on dev2 (Linux, i5-8500, 4 cores), with debug builds and -D warnings as in CI.

Which timeout holds. I ran the flaky test with --exact at 14 nodes, with its CPU capped by systemd-run -p CPUQuota:

CPU cap	5 s (main)	10 s	15 s	20 s	25 s
1.5 cores	3/3 fail (1 at the upload, 2 at the re-estimate after it)	3/3 pass	3/3	3/3	3/3
1 core	3/3 fail (upload)	3/3 pass	2/2	2/2	2/2

Run time doesn't depend on the value: about 210 s at 1.5 cores and about 272 s at 1 core. 10 s is the lowest value tried that held.

What it costs. Full targets at full CPU:

Target	main (5 s, serial teardown)	10 s, serial	this PR (10 s, concurrent)
e2e_chunk (13 tests)	499 s, of which teardown 72 s	504 s / 78 s	437 s / 13 s
e2e_pointer (17 tests, silenced-peer cases)	1149 s / 97 s	not run	1069 s / 15 s

All runs passed. The suites get faster overall, and e2e_pointer's silenced-peer tests are no slower at 10 s.

Round 2 (port 0), on dev2 at ac94cd2:

• cargo clippy -p ant-core --all-targets --all-features -- -D warnings: clean. cargo fmt --check: clean.
• CI's E2E command (--test-threads=1, all 9 binaries): 69 passed, 0 failed: adr0004 3, chunk 13, cost_estimate 7, data 7, file 10, manifest 7, native_payment 4, payment 8, security 10. No "create P2P node" in the log.
• e2e_merkle (35 nodes): 8 passed, 0 failed (2023 s).

CI on round 2 (ac94cd2, run 42012971431): everything green except the inherited Security Audit (RUSTSEC-2026-0387, V2-1743). E2E 69/69 on ubuntu and macOS, Merkle E2E 8/8 on both. No "create P2P node" or panic in any E2E or Merkle log.

CI on round 1 (dec642c): macOS E2E passed e2e_cost_estimate 7/7, with no "need 7"; flake 1 held. The macOS E2E job still failed, on flake 2 (e2e_payment), which round 2 fixes.

New dependency

none. futures is already a dependency of ant-core.

ADR

n/a

Mitigation / rollback

Revert. If a test that waits out a dead peer turns out too slow, lower TEST_CORE_CONNECTION_TIMEOUT_SECS. Don't go to 5 s, which is known to fail.

🤖 Generated with Claude Code

[MaidSafe-Hermes]

Focused round-2 re-review of dec642c957cfa8858e38fc57ff027fd43c4e0927..ac94cd277fd6e38b927d5d7e4d33c0a4da9d3191 (current head ac94cd277fd6e38b927d5d7e4d33c0a4da9d3191). No code findings.

The port-0 change addresses the fixed-port collision. I verified against the pinned saorsa-core source (c3c4c81) that the transport records the actual bound socket address, then exposes the actual listener addresses. With local mode and IPv4-only enabled, those are loopback addresses, so later nodes receive the OS-assigned, usable bootstrap port. Removing retries is reasonable for the diagnosed collision: the kernel now chooses an available port at bind time.

Review coverage: 2/2 seats — sol-1 (gpt-6.1-sol, xhigh) returned UNCERTAIN because its evidence bundle omitted the transport implementation; I resolved that uncertainty by inspecting the exact pinned source. glm-1 (z-ai/glm-5.3) returned APPROVE. The seats did not independently agree; the final disposition follows the source verification.

The required linear-link and pr-template checks pass. E2E/Merkle and other CI jobs are still pending; Security Audit is red for the inherited RUSTSEC-2026-0387 / V2-1743 finding, which is not a required check. This approval is a code-review verdict, not a merge-readiness claim.

[MaidSafe-Hermes]

APPROVE — no blocking code findings on ac94cd2.

I checked the port-0 bootstrap path against the exact saorsa-core revision pinned by Cargo.lock (c3c4c81e63db7e3048f488f7cf0fec4cff3131d4): P2PNode::listen_addrs() returns runtime transport listener addresses, the transport records the bound socket addresses, and local mode uses loopback. The later nodes therefore bootstrap from the actual OS-assigned ports. DEFAULT_NODE_COUNT is 2 × CLOSE_GROUP_SIZE (14); CI exercises the changed test harness and reports E2E and Merkle E2E passing on Ubuntu and macOS.

Review coverage: 2/2 seats reviewed — sol-1 (gpt-6.1-sol, xhigh) returned UNCERTAIN; glm-1 (z-ai/glm-5.3) returned APPROVE. The seats did not agree. I resolved the material uncertainty in sol-1’s report by checking the pinned source, constant definition and CI evidence; this verdict is not presented as seat agreement.

Non-blocking nit: e2e_cost_estimate.rs still says “Use 10 nodes” although DEFAULT_NODE_COUNT is 14. The Security Audit check is failing for the inherited RUSTSEC-2026-0387 / V2-1743 issue noted in the PR; this code approval is not a merge-readiness claim.