[shikhar]

Fixes #196.

cargo test --lib failed on every run in simulation_rescue_preserves_availability_and_work_under_overload. That test (s3_express_overload_sdk_retries) passes when run alone, under --test-threads=1, and under nextest, which CI uses.

Cause

It isn't task scheduling order, as the issue suggests. Each test has its own paused runtime. The shared state is the AWS SDK's retry token bucket: aws-smithy-runtime keeps one in a process-wide StaticPartitionMap per retry partition. The partition defaults to s3-{region}, and every test client uses us-east-1. Simulations running in parallel in one process drew down the same retry quota, so SDK retries were denied depending on what else was running.

Fix

The simulator, the downloader's scripted test clients, and the service's mock client each take a fresh RetryPartition, so every test client gets its own quota, as a separate process would. Production is unchanged: the server builds one client, and sharing the quota across clients is the SDK's intended behavior.

The issue's suggested change, gating the attempt budget on overload alone, would mask the symptom without removing the shared state.

Validation

• cargo test --locked --lib: 190 passed on three runs in a row, where main failed all three.
• nextest: 204 passed. Clippy and fmt are clean.
• The full simulation campaign is identical to main's. Sequential runs within one campaign never drew the bucket down enough to matter, so no reference results change.

🤖 Generated with Claude Code

[greptile-apps]

Confidence Score: 5/5

[Low impact] The PR appears safe to merge; no actionable issues were found.

Summary

Gives each test client a separate AWS SDK retry quota so parallel tests do not consume each other's retries.

• Parallel tests keep separate AWS SDK retry quotas.

_{Reviews (1) · Last reviewed commit: "test: give each test client its own SDK ..." · Reviewed by Greptile}