Reuse the installed run's --wandb-run-id instead of minting one per launch

Every launch generated a fresh wandb run id and wrote it into the
orchestrator argv, even when the release was already installed. Every
leaf payload carries WandbConfig, so a relaunch of an unchanged run
changed the orchestrator command and all leaf payloads, and the
manifest diff rejected the upgrade.

Resolve the id the same way as the run uuid: an id given on the command
line wins; otherwise the installed release's orchestrator command is
read for --wandb-run-id; only a fresh install generates a new one. The
parsed args carry the same id as the argv, so the payloads computed
from them match the pod command.

miles/utils/external_utils/command_utils/helm_backend/launcher/entrypoint.py

@@ -99,7 +99,14 @@ def execute_train(
     installed_manifest = guard.get_manifest(release, namespace)
     run_uuid = _resolve_run_uuid(config, installed_manifest=installed_manifest, release=release)
     env = train_env_vars(request, {}, config=config)
-    pod_argv, args = _compute_train_argv(request, run_uuid=run_uuid, release=release, namespace=namespace, env=env)
+    pod_argv, args = _compute_train_argv(
+        request,
+        run_uuid=run_uuid,
+        installed_manifest=installed_manifest,
+        release=release,
+        namespace=namespace,
+        env=env,
+    )
     deploy_component = DeployComponent(args.deploy_component)
     assert (deploy_component, args.deploy_instance_id) == (config.deploy_component, config.deploy_instance_id), (
         f"the run's pods are told {deploy_component.value}/{args.deploy_instance_id!r}, the release is named "
@@ -312,8 +319,33 @@ def _resolve_run_uuid(config: ExecuteTrainConfig, *, installed_manifest: Manifes
     return generate_run_uuid()
 
 
+def _resolve_wandb_run_id(args: Any, *, installed_manifest: Manifest | None, release: str) -> str | None:
+    if not args.use_wandb:
+        return None
+
+    if (given := args.wandb_run_id) is not None:
+        return given
+
+    if installed_manifest is not None:
+        installed = installed_manifest.flag_value(
+            _WANDB_RUN_ID_FLAG,
+            stateful_set=RunNames.orchestrator_object(release=release),
+            container=naming.ORCHESTRATOR_COMPONENT,
+        )
+        if installed is not None:
+            return installed
+
+    return _generate_wandb_run_id()
+
+
 def _compute_train_argv(
-    request: ExecuteTrainRequest, *, run_uuid: str, release: str, namespace: str, env: dict[str, str]
+    request: ExecuteTrainRequest,
+    *,
+    run_uuid: str,
+    installed_manifest: Manifest | None,
+    release: str,
+    namespace: str,
+    env: dict[str, str],
 ) -> tuple[list[str], Any]:
     argv = [*shlex.split(shell_safe_model_args(request.megatron_model_type)), *shlex.split(request.train_args)]
     assert not ArgvManipulator.is_defined(argv, _ENV_REPORT_FLAG), (
@@ -329,10 +361,10 @@ def _compute_train_argv(
         args = parse_args()
     assert LAUNCHER_REPORT_ENV_VAR not in args.train_env_vars
 
-    # TODO: remove after args refactor handles wandb ids
-    if args.use_wandb and args.wandb_run_id is None:
-        args.wandb_run_id = _generate_wandb_run_id()
-        argv = ArgvManipulator.set(argv, _WANDB_RUN_ID_FLAG, args.wandb_run_id)
+    wandb_run_id = _resolve_wandb_run_id(args, installed_manifest=installed_manifest, release=release)
+    if wandb_run_id is not None:
+        args.wandb_run_id = wandb_run_id
+        argv = ArgvManipulator.set(argv, _WANDB_RUN_ID_FLAG, wandb_run_id)
 
     pod_argv = MooncakeInfo.with_cluster_master(
         argv, plan=_compute_mooncake_plan(args), host=MooncakeInfo.master_service_host(release, namespace)