Event life-history and regime grammar: design boundary
This note records deliberately deferred long-record operations. Names here are design candidates, not implemented public APIs.
Event-relative trajectories
An eventual align_events or event_trajectory operation must require an
anchor (start, peak, or end), pre/post window, overlap policy when events
share observations, spatial extraction policy, and temporal-support alignment
choice. Its result should preserve absolute time, event ID, event_day, anchor
definition, source support, and missing-window behavior. Implementing only an
index shift would conceal scientific choices, so this pass does not expose the
verb.
Period signatures and trends
event_metrics(period="year") already returns unit-labeled ordinary xarray
variables, one row per year. A later period_signature should assemble several
compatible metric objects only after feature names, missing periods, mixed
units, and source scopes have a stable contract.
A future linear trend verb would need an explicit independent coordinate,
calendar-unit conversion, missing-data policy, slope units, uncertainty model,
and autocorrelation caveat. It should not be a thin wrapper that presents an
ordinary regression as a universally valid environmental trend.
Change points and event types
A change-point operation must state its supported methods, null statement, uncertainty, minimum segment length, and representation of multiple candidate changes. No arbitrary algorithm is promoted here.
Event classification should begin with inspectable features: duration, onset rate, decay, severity, footprint, participation, trajectories, and ordering of coupled variables. Opaque clustering remains downstream until those features and regional episode semantics are mature.
Plotting
Event catalogs and metrics remain ordinary pandas/xarray objects and can use their native plotting methods. Specialized defaults for timelines, lag curves, and trajectories are deferred until the underlying scientific objects settle; the plotting layer must not define event meaning.