Make lifecycle tests meaningful and isolate test storage

[jeregrine]

Problem

Several lifecycle scenarios can pass without actually recovering a server: they inspect unchanged metadata, use non-restartable fixtures, or make assertions against unrelated objects. The discovery waiter consumes a timeout rather than observing completion.

Every invocation also contacts LocalStack and clears a shared bucket, even when the selected tests need no storage. Concurrent runs can erase each other's state or collide on EKV directories and node names.

Fix

• Assert recovered process identity, stored state and ownership, and exact restart outcomes. Coordinate concurrent claims at the conditional-write boundary so the losing claim is exercised deliberately.
• Wait for the started discovery task's successful result to be processed, distinguishing idle, failed, and completed cycles. Use a scoped GenServer debug hook rather than adding a production test API; manually driven fixtures disable automatic startup sweeps.
• Make LocalStack and local EKV separate opt-in categories. Lazily create a suite-owned bucket and delete only that bucket after test supervisors stop; remove shared startup wiping and unused IAM setup.
• Allocate worktree-local EKV directories and unique peer names, clean up directories on exit, and restore placement environment variables even after failed assertions.
• Separate scripted retry outcomes from wall-clock deadlines, retain explicit deadline coverage, and wait for reported capacity to be released rather than treating child death as a capacity barrier. Share the bounded eventual-assertion helper.
• Document suite selection, resource isolation, and a pinned LocalStack version.

Supporting information

Only tests, support code, and documentation change; production behavior is unchanged.

The default selection excludes LocalStack, EKV, credentialed integration, and stress cases. Service-backed CI, including #28, must explicitly opt into the storage categories to retain that coverage.

Discovery synchronization overlaps #30, and storage isolation overlaps #33. These are alternative implementations, not independent changes to merge blindly. Unlike #33, this branch also makes local EKV opt-in. #31 adds complementary recovery assertions in the same files; #32 separately strengthens EKV teardown and dirty-state persistence coverage.