Handle missing forecasts before summarisation

[seabbs-bot]

Summary

Closes #1122

• Adds filter_missing_scores() with strategy pattern for dropping incomplete targets
• Adds impute_missing_scores() with strategy pattern for filling missing score rows
• Filter strategies: filter_to_intersection(min_coverage, models)
• Impute strategies: impute_worst_score(), impute_mean_score(), impute_na_score(), impute_model_score(model)
• Shared internal build_missing_grid() detects missing model × target combinations
• .imputed column added to protected columns for tracking

Test plan

• 21 filter tests pass (coverage thresholds, model subsetting, class preservation)
• 35 impute tests pass (all strategies, custom functions, integration with summarise_scores())
• Full test suite (788 tests) passes with no regressions
• Integration pipeline: filter_missing_scores(strategy = filter_to_intersection(models = "baseline")) |> impute_missing_scores(strategy = impute_model_score("baseline"))

This was opened by a bot. Please ping @seabbs for any questions. # checks-passed

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.15%. Comparing base (4a222da) to head (1cf4beb).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1156      +/-   ##
==========================================
+ Coverage   97.98%   98.15%   +0.16%     
==========================================
  Files          37       40       +3     
  Lines        1984     2163     +179     
==========================================
+ Hits         1944     2123     +179     
  Misses         40       40              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[seabbs-bot]

Review

Design feedback (from @seabbs)

Should this even be filter_missing_scores or might people want to filter for other reasons, with the intersection being a special case? Not sure the filter even needs to know the missing grid.

Agreed. Renamed filter_missing_scores() → filter_scores() — a general-purpose filtering function where filter_to_intersection() is just one strategy. Removed the redundant build_missing_grid() call from the filter path entirely. The strategy handles everything independently.

Does the filter always need to do the expand grid thing?

No. filter_to_intersection() works purely by counting coverage per target combination. It never calls build_missing_grid(). The expand-grid detection is only needed on the impute side.

Code review findings (addressed)

1. compare hardcoded as "model" in all impute strategies — impute_worst_score(), impute_mean_score(), impute_model_score() all did setdiff(fu, "model") instead of using the actual compare value. The strategy signature didn't receive compare at all. Fixed: strategy signature changed to function(scores, missing_rows, metrics, compare), all hardcoded "model" references replaced.

2. build_missing_grid() called redundantly in filter — result stored but never passed to strategy. Fixed: removed entirely from filter path.

3. No validation that compare column exists — both functions now have assert_subset(compare, names(scores)).

4. Inconsistent @keywords — all functions now use @keywords handle-metrics.

5. Weak test for impute_worst_score() — was checking global max, now verifies per-target-combination maximum.

This was opened by a bot. Please ping @seabbs for any questions.

[nikosbosse]

I'll wait until you tag me for review, right?

[seabbs]

I think this all works as expected. I was a bit uncertain that compare was the correct name but I think its the best option overall.

[seabbs]

Snapshots look like a ggplot2 change