[adi-kumo]

Problem

structured-data-nims !189 and !197 make the SDM backends (the Kumo Relational and Kumo Tabular NIMs, and the Databricks MLflow model that reuses them) refuse a non-empty inference.inference_config with 400 INVALID_REQUEST. The estimator count is now fixed when the endpoint starts, through SDM_NUM_ESTIMATORS.

This client sent the engine's default config on every KumoRelational request ({kind, num_estimators: 1, column_shuffle: false, ...}), because RelationalAdapter._inference forwarded every key whether or not the caller had set it. So every relational request failed on those backends, although none asked for anything beyond the defaults.

Fix

RelationalAdapter._inference now forwards only the inference-config keys the caller set and omits inference_config when none are left. run_mode, use_prediction_time and output.prediction_statistic are unchanged, and sessions use the same path. An option the caller does set, such as num_estimators: 4, is still sent on its own; a legacy backend honours it and an SDM backend refuses it.

This is safe for the legacy kumo-ml backend: it reads inference_config or {} and applies the rest on top of InferenceConfig.from_task_type(...), whose defaults are identical to the engine's (one estimator, shuffles off, and quantile transforms with a median output for regression). A request without the field gets exactly what this client used to send.

The two Databricks relational notebooks stop passing ENSEMBLE = {'num_estimators': 4, 'column_shuffle': True}, which SDM backends refuse; there the ensemble size is set per endpoint through SDM_NUM_ESTIMATORS. tests/golden/wire_payloads.json is regenerated with scripts/snapshot_wire_payloads.py --write, and it only loses the default config from its one relational case.

Verification

Updated adapter, session and median tests cover a default request sending no config (through both transports), the engine's full defaults not being sent, and an explicit option being sent alone. They fail on main. The client suite passes (1913 passed), along with ruff, mypy and the golden snapshot check.

End to end, an AdventureWorks PREDICT SUM(sales_order_details.OrderQty, 0, 30, days) FOR EACH products.ProductID for product 712 through SDMClient.for_databricks_serving against the SDM staging endpoint kumo-rfm-2-1-stg-v23-mr149-review returns a prediction (120.97) with no inference_config in the request body, while the same request from main gets 400 INVALID_REQUEST.

[Manushpm8]

Verified the central claim against a live relational NIM rather than reading it, and it holds. One defect inline, plus a stale sentence.

The wire loses the config and the answer does not change. Same request, same data, both branches:

	wire inference	prediction
main	run_mode, use_prediction_time, inference_config: {kind, num_estimators: 1, column_shuffle, category_shuffle, hop_shuffle}	[239.27684, 239.424347, 239.129333]
this PR	run_mode, use_prediction_time	[239.27684, 239.424347, 239.129333]

Byte identical, which is what the legacy backend safety argument needed.

A caller set option still travels, alone. inference_config={'num_estimators': 4} puts exactly {"num_estimators": 4} on the wire, and the SDM backend refuses it by name (400 num_estimators is fixed by the selected profile) rather than this client swallowing it. prediction_statistic='mean' still goes on output and changes the answer, so the median default is untouched.

Suites 2691 passed against both live NIMs, lint and hygiene green. The median contract is still pinned: output.prediction_statistic == 'median' is asserted, and the changed assertion correctly stops treating the engine's output_type as the contract.

Found by Codex, confirmed here by running it.

One stale sentence outside the diff: test_decomposition_is_faithful.py:255 still says the AdventureWorks notebook asks with an ensemble, which this PR removes from both notebooks. The test is unaffected, only the sentence explaining why it exists.

[adi-kumo]

Good catch, also fixed in d0abbea. The docstring of test_a_forecast_step_is_the_window_question_asked_on_its_own no longer says the AdventureWorks notebook asks with an ensemble. It now says that an inference config the caller set, here an ensemble, has to reach every step. The test itself is unchanged and still passes num_estimators=4, column_shuffle=True to check that a caller's config reaches every forecast step.

[Manushpm8]

Verified on d0abbea. Editing the config after materialize() no longer changes what is sent, for both the clear and the edit, and the dataclass path holds the same way: mutating a RegressionInferenceConfig after materialize still sends the gathered value, and a bare one sends nothing. Full client suite passes, and against a live relational NIM the no-config, dict-default and dataclass-default forms all return the same prediction.