1200: Fix predicted_label handling and n > 1 column check in categorical metrics

[nikosbosse]

Description

This PR closes #1200.

Three related bugs in the categorical (nominal/ordinal) metrics when called directly with a predicted_label that is not in factor level order (forecasts scored via score() were unaffected, as the pipeline sorts predictions into level order before calling the metrics):

• rps_ordinal() reordered the columns of predicted with the forward permutation predicted[, as.numeric(predicted_label)] instead of the inverse predicted[, order(as.numeric(predicted_label))], so scores were not invariant to how the columns were labelled.
• logs_categorical() ignored predicted_label entirely and indexed predicted[cbind(1:n, as.numeric(observed))], returning the negative log of the probability of the wrong outcome for permuted labels.
• assert_input_categorical() did not check the number of columns of predicted in the n > 1 branch (the n == 1 branch already did), so e.g. a 3x4 matrix with rows summing to 1 was silently accepted and scored meaninglessly.

The fix reorders predicted with order(as.numeric(predicted_label)) in both metrics (preserving NA-observed and single-observation behaviour) and adds ncols = N to the n > 1 assert_matrix() call. Note the latter is a new error on previously (silently, wrongly) accepted input. An existing test expectation in test-metrics-ordinal.R encoded the wrong permutation and has been corrected; permutation-invariance tests and a wrong-width-matrix error test have been added.

Tests were written first and failed against the unfixed code:

FAILURE: 'test-metrics-nominal.R:107:3' -----------
Expected `assert_input_nominal(observed, predicted_wide, predicted_label)` to throw a error.

FAILURE: 'test-metrics-nominal.R:162:3' -----------
Expected `res` to equal `logs_categorical(observed, predicted, predicted_label)`.
Differences:
  `actual`: 2.30 0.36 1.61
`expected`: 0.22 1.61 0.92

FAILURE: 'test-metrics-ordinal.R:138:3' -----------
Expected `rps_ordinal(observed, predicted[, perm], predicted_label2)` to equal `rps_ordinal(observed, predicted, predicted_label)`.
Differences:
  `actual`: 0.820000000 0.530000000 0.200000000
`expected`: 0.050000000 0.500000000 0.200000000

After the fix, the affected test files and the full test suite pass (0 failed, 910 passed).

Dev note: bug identified by an LLM audit (#1189); fix and tests implemented with LLM support, directed by @nikosbosse.

Checklist

• My PR is based on a package issue and I have explicitly linked it.
• I have included the target issue or issues in the PR title as follows: issue-number: PR title
• I have tested my changes locally.
• I have added or updated unit tests where necessary.
• I have updated the documentation if required. (N/A — no roxygen/documentation changes were needed; the documented behaviour is unchanged.)
• I have built the package locally and run rebuilt docs using roxygen2. (N/A — no roxygen comments were changed, so no man/ files needed rebuilding.)
• My code follows the established coding standards and I have run lintr::lint_package() to check for style issues introduced by my changes.
• I have added a news item linked to this PR.
• I have reviewed CI checks for this PR and addressed them as far as I am able.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.15%. Comparing base (06ed56d) to head (26da703).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1203   +/-   ##
=======================================
  Coverage   98.15%   98.15%           
=======================================
  Files          41       41           
  Lines        2225     2226    +1     
=======================================
+ Hits         2184     2185    +1     
  Misses         41       41           

☔ View full report in Codecov by Harness.
📢 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]

Automated review (Claude Fable, directed by @nikosbosse):

Verdict: approve. All checks passed; no must-fix findings.

• Fail-before/pass-after verified independently: on this branch the changed test files pass; with R/ reverted to origin/main (tests kept) the new/corrected expectations fail exactly as claimed — missing ncols error for the 3x4 matrix, and wrong-permutation scores for both logs_categorical() and rps_ordinal() (5 failures total, matching the PR description).
• Full suite green: 0 failed, 910 passed (R_DATATABLE_NUM_THREADS=2).
• Fix correctness: order(as.numeric(predicted_label)) is the correct inverse permutation; drop = FALSE preserves the n = 1 path. Independent edge-case checks all correct: n = 1 with vector input and permuted labels (both metrics), NA in observed under permuted labels, 2-level ordinal with reversed labels, unused observed levels, and wrong-width matrix via the ordinal path. The fix's assumption that predicted_label is a full permutation of the levels is guarded by pre-existing validation (len = N, any.missing = FALSE, empty.levels.ok = FALSE, set-equal levels), so duplicated/partial labels are rejected before reordering.
• Tests are meaningful: the invariance tests use a 3-cycle permutation, which distinguishes the forward from the inverse permutation (a transposition would not); the corrected pre-existing expectation (predicted[, order(c(3, 1, 2))]) is verifiably right, and the nominal test additionally pins -log(p_observed) explicitly.
• Scope/conventions: diff limited to the two metric files, the two test files, and NEWS; NEWS entry accurate (including the new-error caveat per the maintainer decision); title format correct; no lint issues on changed lines; no roxygen changes needed (documented nxN contract unchanged — the stricter check now enforces it).

[nikosbosse]

lgtm