0.2.0: calibrated intervals by default, with the served method reported

[mstrathman]

Problem

The default prediction interval is the in-sample Gaussian band, which we have measured as overconfident on smooth series (57% coverage at a nominal 90%; see the benchmarks page). The calibrated alternative (interval_method: "conformal") is opt-in, restricted to the statistical models, and refuses on series too short to calibrate. A product whose contract is "fail loudly, never serve a flattering number" should not serve a flattering band by default.

Proposal

Make the interval method adaptive by default, with the served method always visible:

• Use conformal when the series and the serving model allow it (enough history for the calibration folds; statistical models).
• Fall back to the residual band otherwise, and report which method served in the result document as a new "interval" field ("conformal" | "residual"), so the fallback is visible rather than silent.
• interval_method stays as an explicit override in both directions.

Since auto model selection already runs rolling-origin machinery in the default path, the marginal cost of default calibration is far smaller than when the current default was chosen.

Scope

• Serving change in the aggregate path plus the new document field.
• forecast_rows() unchanged (the field rides in the document).
• Tests: default-path coverage assertions, short-series fallback visibility, override behavior.
• Benchmarks: re-verify the coverage table; update the docs claim from "the conformal option exists" to "calibrated by default where history allows".
• Changelog entry under the 0.2.0 section; this is the headline change for that release.