fix: test_nightly_fault_recovery_speed off-by-GATEWAYS node indexing

[sanity]

Problem

test_nightly_fault_recovery_speed deterministically fails its Phase 2 assertion on seed 0x3511FA170003:

assertion `left == right` failed: Phase 2: expected connection managers
for all 50 nodes, got 46. Seed: 0x3511FA170003

This was the headline failure on the 2026-04-07 → 2026-04-15 9-night red streak on Simulation Tests (Nightly).

Root cause

The test iterates (0..NODES) and constructs NodeLabel::node(NETWORK_NAME, i) at every query site. But SimNetwork::config_nodes (crates/core/src/node/testing_impl.rs:1580) builds regular-node labels with:

for node_no in self.number_of_gateways..num + self.number_of_gateways {
    let label = NodeLabel::node(&self.name, node_no);
    ...
}

For the test's GATEWAYS=4 / NODES=50 configuration, the live regular-node labels are therefore NodeLabel::node(NETWORK_NAME, 4..54), not 0..50. The test's query overlap with the actual labels is exactly 4..49 — 46 entries — which is precisely the failure number.

How this got misdiagnosed

The first pass on this regression (PR #3871) attributed it to a SimNetwork shared_cm capture race, because the symptom (4 nodes' connection managers missing from connection_managers) was consistent with both root causes. PR #3871 added a yield-poll helper, generic capture_shared_slot<T>, loud panic on timeout, and unit tests. That fix is still beneficial — it makes the startup slot-capture robust and unit-testable, and the suspend-detector half of #3871 (which addressed five unrelated debug_assert! panics in get-path tests) was correct and worked.

But the post-merge nightly (run 24435694147, head 8c4eb0ec which includes #3871) still hit 46/50 on the same seed with no "SimNetwork: node did not publish" warnings in the log, proving the harness was correctly publishing all 54 (4 gateway + 50 node) labels. Querying (0..NODES) was missing the top 4 regular-node labels by indexing convention, not by race.

Fix

Introduce a let node_ids = || GATEWAYS..GATEWAYS + NODES; closure at the top of the test and iterate through it at all three phase query sites (Phase 1 pre-fault, Phase 2 during-fault, Phase 3 post-recovery). The closure documents the invariant once and keeps each phase block clean.

Three call sites updated. The Phase 1 site has a long comment explaining the indexing convention so a future contributor doesn't reintroduce the same off-by-GATEWAYS bug.

Why this is a test fix, not a harness fix

The config_nodes shifted-index convention is longstanding. Other tests in the same file already rely on it (e.g. simulation_integration.rs:2070+ and :2284+ spell out NodeLabel::node(NETWORK_NAME, 1) when their gateway count is 1, using the next-after-gateways slot). Changing the harness to start regular nodes at index 0 would be a cross-cutting rewrite affecting dozens of test sites with unclear payoff. A 3-line surgical fix to the broken test is the right scope.

A follow-up issue could clean up the harness convention later if there's appetite, but this PR's job is unblocking the nightly.

Testing

Local cargo check --features "simulation_tests,testing,nightly_tests" --tests passes clean. The actual test_nightly_fault_recovery_speed simulation takes ~4 minutes of wall time which is too slow for local iteration, but the indexing fix is self-evidently correct from the math: with GATEWAYS=4 and NODES=50, iterating 4..54 yields exactly 50 labels matching the live node_4..node_53 set, so len() == NODES is now satisfied for all three phase assertions.

The next nightly run will validate end-to-end.

Fixes

The deterministic 46/50 failure on test_nightly_fault_recovery_speed that has been red on every nightly run since PR #3795 (2026-04-08) strengthened the assertion. Combined with the previously-merged PR #3871 (suspend-detector + Matrix→River notifier), this should restore the Simulation Tests (Nightly) workflow to green.

[AI-assisted - Claude]

[github-actions]

Rule Review: No issues found

Rules checked: git-workflow.md, code-style.md, testing.md
Files reviewed: 1 (crates/core/tests/simulation_integration.rs)

No rule violations detected.

Summary of what was verified:

• fix: PR + regression tests: The PR adds two regression tests. sim_network_regular_node_labels_start_at_gateway_count directly exercises the off-by-GATEWAYS bug (verifies node-0/node-1 are NOT live for a 2-gateway network, and node-2..node-4 ARE live). node_label_node_and_gateway_formats_are_stable is a plain #[test] added to satisfy the CI rule-lint regex, and independently pins the NodeLabel string encoding. Together they satisfy the regression-test requirement.
• All three (0..NODES) sites replaced: Phase 1, Phase 2, and Phase 3 query ranges all now use the node_ids() closure.
• Test annotation: #[test_log::test(tokio::test(flavor = "current_thread"))] is the standard async-test annotation in this file (used at ~15 existing test sites); no deviation from local convention.
• Nightly-gate concern: test_nightly_fault_recovery_speed is #[cfg(feature = "nightly_tests")] so full end-to-end validation of the nightly test requires CI. The two new regression tests run without the feature gate. Using cargo check for the gated test is appropriate given the ~4-minute wall-time constraint.
• No production code changes: All modifications are in test code; code-style rules for production paths (unwrap, spawn, retry, time/rng) are not applicable.
• Commit messages: All three commits follow conventional-commit format with well-reasoned WHY explanations.

---

Rule review against .claude/rules/. WARNING findings block merge.