[NewGraphEnvironment]

Summary

• cd_anomaly() resolves anomaly_type, unit (the anomaly's unit, same meaning as cd_variables()$unit) and long_name from the input where present, falling back to cd_variables() row by row. A variable whose type can't be resolved is now an error naming it, not a silent NA. Contract documented under "Input contract" in ?cd_anomaly, with a runnable non-ERA5 example.
• cd_trend() carries long_name through, plus anomaly_type/unit when trending anomalies (not raw values, where the anomaly unit would mislabel the slope). cd_summary() uses them, then the registry, then the variable name.
• One row per (variable, period, year) is enforced in cd_baseline, cd_anomaly, cd_compare, cd_trend, so two stations stacked under one variable are refused instead of pooled. Grouped input is accepted. The rules live in three shared helpers in R/cd_anomaly.R (series_check, meta_resolve, meta_check).

Behaviour changes for existing callers

• Input that used to return NA anomalies now errors.
• Input with duplicate (variable, period, year) rows used to be pooled silently; it now errors. All committed vignette series have 0 duplicates.
• cd_trend(cd_anomaly(...)) output has two extra columns (anomaly_type, unit).

Suggest a minor bump at merge.

Related Issues

• Fixes cd_anomaly() and cd_summary(): accept series outside cd_variables() instead of returning NA #92
• Unblocks Flow in date windows at hydrometric stations: daily series, window metrics, departure via cd wet#25
• Follow-up: cd_plot_timeseries() and cd_plot_comparison(): use long_name/unit carried on the input #93 (plot labels)
• Relates to NewGraphEnvironment/sred-2025-2026#23

Test plan

• devtools::test(): 278 pass, 0 fail (2 warnings are pre-existing, in test-cd_plot_comparison.R)
• 21 mutations, one per new guard and chain-caller check, each turn a test red
• ERA5 path unchanged: the regional Peace and Kootenay cd_trend(ano) reproduce the committed trn exactly, and cd_summary() output is identical
• pkgdown CI. pkgdown::check_pkgdown() fails locally on main as well (DESCRIPTION URL), and no exports were added

Notes

Review: one plan review and three code-check rounds found 10 issues and all were fixed. Two sat inside earlier fixes. The last was a plain ungroup() that turned station-grouped input from an error into silent pooling. The loop ended by enumerating every export against the shared helpers, not on a clean round. Details are in planning/archive/2026-09-issue-92-series-outside-cd-variables/.

Known and not fixed, both pre-existing: cd_trend() on zero rows returns a tibble with no columns, and a series with no baseline row still gets anomaly = NA.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GignjoKpZADXywTf6pAMgj