1198: Fix bias_quantile() quantile level sorting and na.rm crash

[nikosbosse]

Description

This PR closes #1198.

bias_quantile() had two related bugs in bias_quantile_single_vector() (R/metrics-quantile.R):

1. Wrong results with unsorted quantile levels: predictions were reordered by quantile level (predicted <- predicted[order]) but quantile_level itself was left unsorted, so predictions and levels became mispaired. The same forecast then produced different bias values depending on input order, e.g. bias_quantile(1.5, c(1, 2, 3), c(0.25, 0.5, 0.75)) returned 0.5 while the identical forecast bias_quantile(1.5, c(3, 1, 2), c(0.75, 0.25, 0.5)) returned -0.5. Fixed by sorting quantile_level together with predicted.

2. Crash with na.rm = TRUE: when NA removal stripped all quantile levels on one side of the median, interpolate_median() produced a length-zero median (via max(numeric(0)) returning -Inf with a warning) and the function crashed with argument is of length zero. Fixed by returning NA_real_ when the median is no longer interpolable after NA removal, consistent with na.rm = FALSE, which also returns NA_real_ silently.

New regression tests were written first and confirmed to fail on main:

Failure ('test-metrics-quantile.R:817:3'): bias_quantile() is invariant to the order of quantile levels
Expected `bias_quantile(...)` to equal `0.5`. Differences: `actual`: -0.5, `expected`: 0.5

Error ('test-metrics-quantile.R:766:3'): bias_quantile() returns NA when na.rm removes one side of the median
Error: argument is of length zero
Warning: no non-missing arguments to max; returning -Inf

After the fix, the full test suite passes (0 failed, 911 passed). Note one intentional behaviour change beyond bugfixing correctness: internally consistent forecasts passed with unsorted quantile levels, which previously either returned wrong values or errored with "Predictions must not be decreasing", now score correctly; genuinely inconsistent input still errors as before (existing tests unchanged).

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 user-facing documentation changes needed; no roxygen comments were changed.)
• I have built the package locally and run rebuilt docs using roxygen2. (N/A — no roxygen comments were changed, so no doc rebuild was needed.)
• My code follows the established coding standards and I have run lintr on the changed files 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 (dfaa4d6).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1205   +/-   ##
=======================================
  Coverage   98.15%   98.15%           
=======================================
  Files          41       41           
  Lines        2225     2228    +3     
=======================================
+ Hits         2184     2187    +3     
  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. Independently verified fail-before/pass-after and found no defects.

Checks performed:

• Fail-before/pass-after confirmed: with R/ reverted to main (tests kept), the new tests fail exactly as claimed — order-invariance test returns -0.5 instead of 0.5, and the na.rm test errors with "argument is of length zero" plus the max() -Inf warning. On the branch, both pass.
• Full test suite green on the branch (0 failures; 18 environment-dependent skips, mostly vdiffr).
• Edge cases probed beyond the PR's tests, all correct: a 200-case random permutation grid (unsorted levels bit-identical to sorted); all-NA predictions with na.rm = TRUE → NA; NA in quantile_level (not predicted) stripping one side → NA; NA at the 0.5 level with both sides surviving → correctly interpolated; unsorted + NA combined; single 0.5-level input; matrix input with unsorted levels; genuinely inconsistent input still errors.
• Brief compliance: matches maintainer decisions exactly (sort predicted and quantile_level together; silent NA_real_ when NA removal leaves the median non-interpolable, no warning).
• Scope/conventions: diff touches only R/metrics-quantile.R, the test file, and NEWS.md; NEWS entry accurate; PR title format correct; the intentional error-to-value behaviour change for consistent-but-unsorted input is disclosed in the PR body and I confirmed it behaves as described.

No must-fix findings.

[nikosbosse]

lgtm