[992X]

Issue #, if available: N/A

Description of changes:

Refreshes sagemaker-core's codegen input from the public botocore model and
regenerates the layers built from it. The model file is the upstream bytes
verbatim, and the generated layers are the output of
python -m sagemaker.core.tools.codegen.

	before	after
operations	403	409
shapes	3005	3084

What the refresh brings in

This is a whole-model sync, so the diff is not limited to any one feature.

Operations (6 new)

• AttachClusterNodeNetworkInterface
• CreateTrainingSession, DescribeTrainingSession, ListTrainingSessions,
  StopTrainingSession, DeleteTrainingSession

Shapes (79 new, 66 changed)

• 19 of the new shapes belong to the training-session operations; the other 60
  are elsewhere in the API.
• The 66 changed shapes span AIBenchmark*, AIRecommendation*, App* and
  Cluster*, among others.
• NotebookInstance gains the PendingMaintenance and InMaintenance states.

Resources and configuration schema

• A TrainingSession resource, derived from the standard CRUD naming:
  create / get / get_all, and delete / refresh / stop /
  wait_for_delete / wait_for_status.
• MlflowApp gains kms_key_id in the configuration schema.

No caller is switched over to any newly generated resource in this PR. This
change only brings the generated layer back in line with the model.

One codegen change, for a resource whose identifier is nested

DescribeTrainingSession returns everything inside a single
TrainingSessionDetail member, so TrainingSessionName is not hoisted onto the
resource class. Without the identifier as a class attribute the generated
resource could not complete a lifecycle: stop() raised AttributeError,
refresh() and delete() took the identifier as a required positional while
wait_for_status() / wait_for_delete() called refresh() with none, and
get_name() returned None.

resources_codegen.py already keeps a list of resources whose identifier is
absent from their describe output — ImageVersion, Workteam, Workforce,
SubscribedWorkteam, HubContent — and injects it as a class attribute for
exactly this reason. TrainingSession is added to that list. The nested value
stays reachable at training_session_detail.training_session_name.

After the change, with the boto client stubbed:

model fields: ['training_session_name', 'training_session_detail']
  refresh()   delete()   stop()

get_name()        -> 'my-session'
stop()            -> OK
refresh()         -> OK
delete()          -> OK
wait_for_status() -> OK

wire calls:
  stop_training_session({'TrainingSessionName': 'my-session'})
  describe_training_session({'TrainingSessionName': 'my-session'})
  delete_training_session({'TrainingSessionName': 'my-session'})

The more general fix would be to unwrap a single-member describe response when
collecting class attributes, the way resources_extractor already does when
building status chains. That changes generation for every resource, so it is
left alone here and worth considering separately.

Notes for anyone regenerating

Two properties of the generator are worth writing down, because both produce a
diff that looks plausible but is wrong:

1. codegen needs two passes to converge. generate_resources() reads the
   existing config_schema.py to decide which resources get a
   populate_inputs_decorator, and the schema is regenerated after it. A
   single pass therefore emits new resources without the decorator, and those
   resources silently ignore sagemaker_config defaults. A third pass is a
   no-op, so two passes are enough. Both passes are reflected here.
2. resources_extractor writes resource_plan.csv to the current working
   directory. Running the unit tests from sagemaker-core/ overwrites the
   tracked file with a partial plan, so regenerate before diffing rather than
   after running tests.

Separately, codegen does not emit the module docstrings at the top of
config_schema.py or shape_dag.py. Neither file is excluded from flake8 or
docstyle, so dropping them fails the codestyle gate. Both are preserved here.

Testing

• sagemaker-core unit tests on this branch and on unmodified master: both
  14 failed, 3922 passed, 41 skipped, with identical failing test ids, so the
  regeneration introduces no new failures. The pre-existing failures include
  tests/unit/tools/test_api_coverage.py::test_api_coverage, which fails on
  master as well.
• pydocstyle on src/sagemaker: 0 violations, matching master.
• black at the pinned version: the only file flagged is
  tests/unit/local/test_entities.py, flagged identically on master.

Not included

No VERSION or CHANGELOG.md changes.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[github-actions]

🤖 AI code review

This PR is a whole-model botocore sync of sagemaker-core: it refreshes service-2.json, resource_plan.csv, and the generated resources.py, shapes.py, config_schema.py, and shape_dag.py. Most of the diff is faithfully regenerated and additive (new shapes defined before use, new MLflow/IdC/observability/HyperPod fields, NotebookInstance maintenance states, etc.), consistent with prior sync PRs. However, the newly generated TrainingSession resource is functionally broken because its identifier (TrainingSessionName) lives inside the nested TrainingSessionDetail describe output rather than at the top level, so the generated accessor code is inconsistent and will raise at runtime.

_{Reviewed commit 6f21b06. Automated review; verify before acting.}

[github-actions]

🔴 High: The generated TrainingSession resource is not usable as emitted. DescribeTrainingSessionResponse wraps everything in TrainingSessionDetail, so the only class attribute is training_session_detail — there is no training_session_name or training_session_status attribute on the resource. As a result:

• stop() builds {"TrainingSessionName": self.training_session_name}, but self.training_session_name does not exist on the model (it lives under self.training_session_detail.training_session_name), so calling stop() raises AttributeError.
• refresh(self, training_session_name) and delete(self, training_session_name) take the identifier as a required positional argument, but wait_for_status() and wait_for_delete() call self.refresh() with no arguments (as every other resource's wait methods do), which raises TypeError: refresh() missing 1 required positional argument.
• get_name() looks for a training_session_name attribute and will fail to find it, logging an error and returning None.

This is the generator's handling of a resource whose identifier is nested in a detail object; the fix likely belongs in codegen (e.g. a REFRESH_IDENTIFIER_FALLBACKS-style mapping or chained-attribute resolution to pull training_session_name from training_session_detail) rather than hand-editing the output. Please verify the regenerated TrainingSession resource against the new model before relying on it; as generated it cannot complete a stop/wait/refresh cycle.

[github-actions]

🤖 AI code review

This PR syncs sagemaker-core's bundled SageMaker service model with upstream botocore and regenerates the layers built from it (service-2.json, resource_plan.csv, resources.py, shapes.py, shape_dag.py, config_schema.py), plus one small hand-written change to the generator (resources_codegen.py) adding TrainingSession to the list of resources whose identifier must be injected as a class attribute.

I spot-checked the new TrainingSession resource end to end: the injected training_session_name identifier is used consistently by create/get/refresh/delete/stop, wait_for_status/wait_for_delete read the nested training_session_detail.training_session_status, and all referenced new shapes (TrainingSessionDetail, LoraConfig, TrainingSessionModelConfig/OutputConfig, BillableTokenUsage, etc.) are defined in shapes.py and shape_dag.py. The resources_codegen.py change follows the existing ImageVersion/Workteam/HubContent pattern. Additive parameters (included_data, kms_key_id, idc_config, training_plan_arns, new wait_for_status literals) are backward-compatible, and I saw no required-positional reordering of the kind flagged in prior syncs. No concrete issues found; consistent with precedent that pure codegen syncs need no additional unit tests (Source: #6285).

_{Reviewed commit 837c52c. Automated review; verify before acting.}