[Bug]: creation_policy: reuse ignored #4045

Steps to reproduce

1. Create a fleet:

   dstack fleet -v
    NAME     NODES  RESOURCES  SPOT  BACKEND  PRICE  STATUS  CREATED      ERROR
    backend  0..    -          auto  *        -      active  30 mins ago
   

2. Submit a run with creation_policy: reuse/--reuse:

    Project              main
    User                 dmitry-local-admin
    Type                 dev-environment
    Resources            cpu=1.. mem=1GB.. disk=1GB.. gpu=0..
    Spot policy          on-demand
    Max price            off
    Retry policy         off
    Creation policy      reuse
    Idle duration        5m
    Max duration         off
    Inactivity duration  off
   
   No matching instance offers available. Possible reasons: https://dstack.ai/docs/guides/troubleshooting/#no-offers
   

   Note Creation policy reuse

Actual behaviour

A new instance provisioned

Expected behaviour

The run fails with FAILED_TO_START_DUE_TO_NO_CAPACITY – Could not reuse any instances for this job

dstack version

0.20.28

Server logs

DEBUG    dstack._internal.server.services.backends:492 Requesting instance offers from backends: ['aws', 'runpod', 'gcp']
...
INFO     dstack._internal.server.services.events:205 Emitting event: Job created on run submission. Status: SUBMITTED. Event targets: job(0eaa31)dev-environment-0-0. Actor: system
...
INFO     dstack._internal.server.services.events:205 Emitting event: Instance created for job. Instance status: PENDING. Event targets: instance(2aae42)backend-0, job(0eaa31)dev-environment-0-0. Actor: system
...
DEBUG    dstack._internal.server.background.pipeline_tasks.jobs_submitted:2240 job(0eaa31)dev-environment-0-0: trying t3.micro in aws/us-east-1 for $0.0104 per hour
...
INFO     dstack._internal.server.services.events:205 Emitting event: Job status changed SUBMITTED -> PROVISIONING. Event targets: job(0eaa31)dev-environment-0-0. Actor: system

Additional information

JobSubmittedWorker.process, first pass:

• _process_assignment -> _select_assignment picks a fleet and returns _NewCapacityAssignment
• _apply_assignment_result creates a placeholder instance and sets both JobModel.instance_assigned and JobModel.instance

JobSubmittedWorker.process, second pass:

• _process_provisioning, job_model.instance is set -> _process_new_capacity_provisioning, run.run_spec.merged_profile.creation_policy == CreationPolicy.REUSE check below is unreachable