Verify agent startup resume checkpoints

[asim]

Adds a focused startup-recovery regression for durable agent checkpoints. The test recreates an agent on the same checkpoint store, drains pending runs with ResumePending, and asserts the already-completed tool side effect is not replayed while the run reaches a done checkpoint.

Closes #4341
Closes #4349

Testing:

• go build ./...
• go test ./agent -run TestResumePendingAfterFreshAgentRestartDoesNotReplayCompletedTool -count=1 -timeout=60s
• go test ./agent
• go test ./... (fails in this environment: atlascloud stream conformance attempts a real request; loopback gRPC tests are blocked by the sandbox envoy)
• golangci-lint run ./...

[Copilot]

Pull request overview

Adds a regression test to validate durable agent checkpoint recovery on service startup: recreating an agent against the same checkpoint store and draining work via ResumePending should complete the run without replaying already-completed tool side effects.

Changes:

• Add TestResumePendingAfterFreshAgentRestartDoesNotReplayCompletedTool, which simulates a process stop after a tool call is checkpointed and verifies a fresh agent instance can resume to done.
• Assert ResumePending drains the backlog (no remaining Pending runs) and produces a done run summary without re-executing the tool handler.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.