E2E: back-migration test flakes on Connection refused after restart

[barakeinav1]

Background

e2e-tests::e2e migration_service::migration_service__should_handle_back_migration_a_to_b_to_a (added in #3346) has been flaking on CI. Three confirmed failures with identical signature (Connection refused on http://127.0.0.1:21725/debug/migrations at migration_service.rs:533:29):

• PR test(e2e): back-migration regression test #3346 CI run 26451546911 attempt 1
• PR fix(node): refresh attestation before concluding back-migration (#2121) #3362 CI job 77946818099
• PR fix(node): refresh attestation before concluding back-migration (#2121) #3362 CI job 77952041274

Root cause: cluster.wait_for_node_healthy() only checks that /health returns 200 OK. /health is || async { "OK" } (crates/node/src/web.rs) — bound very early in node startup, before the indexer initializes. After the kill+restart in the back-migration test, the function returns in as little as 16 ms (the time to spawn the binary and bind a socket), with the indexer still warming up. The test then marches into the back-migration's polling loop on /debug/migrations, and if A0's process exits during catch-up (likely panicking on the contract state after the forward migration's identity swap), the polling sits on Connection refused for the full 30 s INDEXER_SYNC_TIMEOUT.

User Story

As a developer, I want migration_service__should_handle_back_migration_a_to_b_to_a to pass deterministically so flaky CI doesn't block unrelated PRs.

Acceptance Criteria

• The test passes on at least 15 consecutive CI runs on the same branch.
• The fix is targeted at the readiness-gap (don't proceed before the indexer is making progress), not a band-aid on the polling timeout.

Resources & Additional Notes

• Fix PR: test(e2e): wait for indexer progress in back-migration restart (fix flaky test) #3365 (validated with 15/15 success and per-run wait metrics).
• If the secondary "A0 crashes during catch-up" theory ever materializes after this fix, the new helper's error message will say "indexer block-height metric not available — node may have exited" (instead of the silent 30 s Connection refused timeout). A future change to capture each MPC node's stderr.log into the test output on failure would close the diagnostic loop for that case.