test: record controller polling delays instead of sleeping them

[alongubkin]

Problem

SingleControllerExecutor (crates/alien-infra/src/core/controller_test.rs) slept for each controller's suggested_delay between steps. Those are 2–30 s, and a create+delete flow has 10–20 polls. The mocked providers answer at once, so this was pure idle time. GCP worker flows took 37–62 s each, and alien-infra spent ~960 test-seconds per CI run.

Nothing checked the delays either. A controller that regressed to polling a cloud API in a hot loop would still pass.

Change

• DelayMode::Record is the default. The executor records each suggested delay (suggested_delays()) and continues.
• .real_delays() keeps real waits for the three #[ignore] live-cloud tests (azure_sandbox_image_live, azure_sandbox_image_refusals_live, kubernetes_sandbox_live).
• A suggested delay over 10 minutes (MAX_SUGGESTED_DELAY) fails the run. That catches a seconds/milliseconds mix-up or a runaway backoff.
• The GCP and AWS worker create+delete flows check each phase separately (take_suggested_delays, assert_polling_delays).
  • Every wait must be at least 1 s.
  • GCP create and delete, and AWS create, must poll at least once.
  • AWS Lambda deletion completes in the call, so its delete phase may not poll.
• Two unit tests for the executor: delays are recorded without sleeping, and an over-limit delay is rejected.

Safe to stop sleeping?

• Steps, mock calls and assertions are unchanged. Only the idle time between steps goes away.
• The Azure wall-clock waits (RBAC, managed identity, Container Apps) are already 0 under #[cfg(test)]. The seeded-timestamp tests (sandbox build timeout, retire window) seed their times explicitly.
• Paused tokio time isn't used: these tests talk to a real local httpmock server, and paused time would let reqwest timeouts fire early.

Verified

• cargo test -p alien-infra --features all-platforms,kubernetes,local,test --lib --tests: 813 lib tests pass in 11 s, plus the integration tests.
• CI benchmark (same harness change, 2 runs): alien-infra 961 → 199/209 test-seconds, all tests pass.

🤖 Generated with Claude Code

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Test infrastructure records polling delays instead of sleeping.

The PR appears safe to merge; no outstanding finding or new actionable issue remains.

Summary

The PR replaces sleeps with recorded polling delays in mocked controller tests, retains real waits for three live tests, and checks worker polling delays separately for create and delete.

• The follow-up change separates phase assertions using take_suggested_delays() and adds a shared minimum-delay assertion.

_{Reviews (2) · Last reviewed commit: "test: check polling delays per phase"}

[alongubkin]

@greptileai review