pre_start hooks: per_replica: true is accepted by the schema but always rejected at runtime

What this issue is about

pre_start hooks support a per_replica attribute in the compose-spec model (types.PreStartHook.PerReplica, compose-go's types/hooks.go): "runs the hook once per service replica instead of once per service." docker/compose's engine, however, does not implement it: runPreStart (pkg/compose/pre_start.go) validates every declared hook up front and rejects the whole service if any hook sets per_replica: true:

for i, hook := range service.PreStart {
    if hook.PerReplica {
        return fmt.Errorf("service %q pre_start[%d]: per_replica is not yet supported; remove per_replica or set it to false", service.Name, i)
    }
}

So today, per_replica: true is accepted by the loader/schema but always a hard error at runtime — there is no path to actually use it.

Illustration

services:
  app:
    image: alpine
    command: sleep infinity
    deploy:
      replicas: 3
    pre_start:
      - image: alpine
        command: ["sh", "-c", "echo initializing replica"]
        per_replica: true

$ docker compose up -d
service "app" pre_start[0]: per_replica is not yet supported; remove per_replica or set it to false

The only way to get this compose file to run at all today is to drop per_replica: true, at which point the hook runs once for the whole service (against the lowest-numbered replica only, per the current one-hook-one-runner model) instead of once per replica — a different, sometimes unacceptable behavior for hooks that need to prepare per-replica state (e.g. per-replica config files, per-replica volume initialization).

Suggested scope for an implementation

The reconciliation-plan work in #14221 (pre_start hook runners as first-class plan resources) planned one CreateHookContainer operation per declared hook, targeted at a single resolved replica. Supporting per_replica: true would mean planning one runner per hook × per replica instead, with:

• a deterministic name/label scheme distinguishing replicas (e.g. extending the existing com.docker.compose.hook-index label with a replica index, or a dedicated label),
• VolumesFrom resolved against that replica rather than the lowest-numbered one,
• the corresponding execution-phase change in runPreStart/execPreStartHook to run and account for one runner per replica.

This issue only tracks the feature gap; the reconciliation-plan groundwork from #14221 already discussed and worked around it (see its planPreStartHookRunners, which explicitly skips planning entirely for a service with any per_replica: true hook, to avoid orphaning runners that could never execute).