Fix duplicate column names from summarise_scores() with empty metrics (#1179)

[nikosbosse]

CLAUDE: Closes #1179.

Summary

• summarise_scores() previously selected which columns to summarise via colnames(scores) %like% paste(metrics, collapse = "|"). When the metrics attribute is character(0) (i.e. every metric in score() warned and returned nothing), this pattern becomes the empty string, which %like% matches against every column — including the by column. The by column was then passed to the summary function, producing the spurious "argument is not numeric or logical" warning and a data.table with a duplicate by column. The duplicate is invisible inside data.table but breaks downstream conversion to tibble.
• Switched to exact column-name matching via intersect(colnames(scores), metrics). This also incidentally fixes a latent issue where a metric named e.g. "wis" would have matched any column whose name contained "wis" (such as "wis_relative_skill").

Both tightenings from #1179 are implemented

The issue suggested two fixes (either or both); this PR does both:

1. summarise_scores() no longer summarises its by columns. With exact-name matching, .SDcols is now intersect(colnames(scores), metrics), so task-ID columns (including the by column) are never passed to the summary function. This removes the duplicate-column root cause.
2. summarise_scores() errors early when there are no score columns to summarise. When intersect(colnames(scores), metrics) is empty there is nothing meaningful to return, so it now aborts with a clear message rather than producing a malformed object.

Tests

• End-to-end regression test for the issue reprex: scoring example_quantile with only interval_coverage_55 warns and produces no score columns, after which summarise_scores() errors.
• A unit-level test of the same empty-metrics case (manually cleared score columns).
• A test guarding against partial-name matching.

Out of scope

The issue also raises whether score() should itself fail (rather than return an empty scores object) when every metric fails. That's a real question but a bigger design call; leaving it for a separate discussion.

Test plan

• Targeted tests pass locally (testthat::test_file("tests/testthat/test-summarise_scores.R") — 16 pass)
• lintr::lint() clean on changed files
• CI green (full R CMD check / covr relied on CI)
• Branch rebased onto current main, which also picks up the multivariate-sample snapshot fix (Fix macOS CI snapshot precision + lint-changed-files workflow #1182) so macOS CI should be green.

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

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.99%. Comparing base (abd733e) to head (d124e69).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1180   +/-   ##
=======================================
  Coverage   97.98%   97.99%           
=======================================
  Files          38       38           
  Lines        2036     2045    +9     
=======================================
+ Hits         1995     2004    +9     
  Misses         41       41           

☔ 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.

[nikosbosse]

CLAUDE: All checks pass except macOS-latest (release), which fails on the same two tests at the same line numbers (test-class-forecast-multivariate-sample.R:343 and :376 in score.forecast_multivariate_sample() creates expected output) on main as well (e.g. run 25738222052). This is pre-existing macOS-specific breakage in the multivariate sample tests, unrelated to this PR.

[seabbs]

#1179 contains a nice reprex test case can we test against it. I wasn't clear if this was from you @nikosbosse or claude. I guess our practice for review is to wait to be tagged but maybe making it clearer in the PR description would be good:?

[seabbs]

This looks good to go otherwise though it would be nice to explicitly state that both the tightenings suggested in #1179 have been implemented (it looks to me that they have)

[seabbs-bot]

Addressed the review feedback:

• Reprex test: added an end-to-end regression test that scores `example_quantile` with only `interval_coverage_55` (the exact reprex from summarize_scores() produces duplicate column names when input has no score columns #1179) — it warns, produces no score columns, and `summarise_scores()` then errors. Kept the unit-level empty-metrics test alongside it.
• Both tightenings from summarize_scores() produces duplicate column names when input has no score columns #1179: confirmed both are implemented and now spelled out in the PR description — (1) `by` columns are no longer summarised (exact-name `.SDcols` matching), and (2) early error when there are no score columns.
• Up to date with main: rebased onto current `main`, which also brings in the multivariate-sample snapshot fix (Fix macOS CI snapshot precision + lint-changed-files workflow #1182), so the previously failing macOS tests should now pass.

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

[seabbs-bot]

Automated review pass (agent quality gate)

No Critical or Important findings. The change is correct and well-scoped.

Observations:

• The fix replaces a fragile regex partial-match (`%like% paste(metrics, collapse = "|")`) with exact-name matching (`intersect(colnames(scores), metrics)`), which also fixes the latent partial-match bug (e.g. "wis" matching "wis_relative_skill"). Good.
• Turning the previously-silent duplicate-column case into an explicit `cli_abort` is the right call and matches the issue intent.
• Line 82 sets `attr(scores, "metrics") <- metrics` (full vector) rather than `metric_cols`. This preserves prior behaviour and is reachable only when `metric_cols` is non-empty, so it is fine; just noting it is intentional.
• `cli` and `cli_abort` are already package deps/imports; the multi-line message string matches the existing codebase convention (cli collapses internal whitespace on render).
• Regression tests cover the empty-metrics error, the end-to-end reprex from summarize_scores() produces duplicate column names when input has no score columns #1179, and the partial-match guard. `scores_quantile` fixture confirmed present in setup.R.
• `lintr::lint()` clean on both changed files locally.

CI is pending; will continue to monitor checks and mergeability.

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

[seabbs]

@seabbs-bot added the explicit reprex so this looks good to me now.