feat(workloads): add agent state

[casey-brooks]

Summary

• add agent_state persistence and proto mapping
• update touch/sweep logic with workload.updated notifications
• reset last_activity_at on starting->running and add sweep config

Testing

• go test ./...
• go vet ./...

Closes #48

[casey-brooks]

Summary

• add agent_state persistence and proto mapping
• update touch/sweep logic with workload.updated notifications
• reset last_activity_at on starting->running and add sweep config

Tests

• go test ./... (passed: all, failed: 0, skipped: 0)
• go vet ./... (no issues)

[noa-lucent]

Requesting changes — core behavior is close, but a couple gaps will break the activity-sweep semantics and acceptance criteria:

1. starting→running must reset last_activity_at for all status update paths

• Today the reset is only triggered via UpdateWorkload’s ResetLastActivity; UpdateWorkloadStatus can also promote to running but never resets.
• This can cause the sweep to immediately mark agent_state=idle right after a workload starts.
• Please centralize in updateWorkload (SQL CASE based on previous status) or compute/pass the reset from UpdateWorkloadStatus too.

2. Tests: assert agent_state is surfaced in API responses

• Acceptance criteria calls out GetWorkload + ListWorkloadsByThread include agent_state.
• Please add assertions (and a GetWorkload success-path test) to lock down DB-string → proto-enum mapping.

3. (Minor) Add a TouchWorkload test for the processing (no-event) path with NotificationsClient enabled.

[casey-brooks]

Summary

• reset last_activity_at when UpdateWorkloadStatus moves starting -> running
• add GetWorkload/ListWorkloadsByThread agent_state assertions
• add TouchWorkload no-publish coverage for processing state

Tests

• go test ./... (passed: all, failed: 0, skipped: 0)
• go vet ./... (no issues)

[noa-lucent]

Re-review: requested changes are addressed.

• UpdateWorkloadStatus now resets last_activity_at on starting→running (consistent with UpdateWorkload), preventing the activity sweep from prematurely marking workloads idle.
• Added assertions/tests to ensure agent_state is returned by ListWorkloadsByThread + GetWorkload (DB-string → proto-enum mapping covered).
• Added TouchWorkload coverage to ensure no workload.updated publish when already processing.