[BUG] predict: forwards y to non-annotators, drops time index in interval/var modes, unbounded horizon #536

Severity: P2 — output quality / consistency. Verified live 2026-08-09.

Three predict issues, all confirmed over the wire.

NB-18 — predict forwards y blindly to non-annotator estimators

predict(estimator_handle=<forecaster>, y_dataset="arrow_head")
  -> {"error": "BaseForecaster.predict() got an unexpected keyword argument 'y'"}

The y_dataset/y_handle args exist for annotators, but they're passed through to any estimator, producing a raw sktime TypeError for forecasters.
Fix: scitype-aware handling — only forward y to estimators whose predict accepts it; otherwise a clear message ("forecasters do not accept y at predict").

NB-21 — interval/variance modes drop the time index

predict keys by period:

{"1961-01": 417.0, "1961-02": 391.0, ...}

but predict_interval / predict_var return bare arrays with no labels:

"intervals": {"sales_0.8_lower": [215.34, 199.33, ...], "sales_0.8_upper": [...]}

The caller can't map interval/variance values back to time points.
Fix: index-key all modes consistently.

NB-22 — predict payload is unbounded

predict(estimator_handle=<fitted>, horizon=1000) returns 1000 inline period-keyed rows (~30 KB+ JSON), no cap or truncation. Any large horizon floods the client context.
Fix: soft cap / truncated preview with a row count, consistent with the response-limit machinery used elsewhere (and the job-result cap added in commit 2b9aa16).

(Catalogued as NB-18 / NB-21 / NB-22 in MCP_TEST_FINDINGS.md.)