1202: Fix interval_coverage() floating point matching and NaN scores for quantile levels 0 and 1

[nikosbosse]

Description

This PR closes #1202.

Two related edge-case bugs in the quantile-based metrics:

1. interval_coverage() errored on seq()-generated quantile levels. The check for the required quantile levels used an exact floating point comparison (necessary_quantiles %in% quantile_level), so e.g. interval_coverage(observed, predicted, seq(0.05, 0.95, 0.05), interval_range = 30) aborted with "the 0.35 and 0.65 quantiles are required" even though those quantiles were present (0.35 %in% seq(0.05, 0.95, 0.05) is FALSE). The fix rounds both sides to 10 decimal places before matching, consistent with the package's existing convention (e.g. in as_forecast_quantile()).

2. wis(), interval_score() and quantile_score(weigh = FALSE) returned NaN for quantile levels 0 and 1. These levels are explicitly permitted by input validation and occur in real hub data ("min"/"max" quantiles), but they form a 100% interval with alpha = 0, and the penalty terms 2 / alpha * (bound - observed) * indicator evaluated to Inf * 0 = NaN. This broke the documented WIS = mean quantile score identity: wis(5, matrix(c(1, 3, 5, 7, 9), nrow = 1), c(0, 0.25, 0.5, 0.75, 1)) returned NaN while quantile_score() returned 0.4. The fix applies the indicator before the 2 / alpha factor and cancels 2 / alpha * alpha / 2 analytically in the weighted case; in the unweighted case, zero penalties stay zero and the score is Inf when the observation falls outside a 100% interval (mathematically correct, per maintainer decision). quantile_score(weigh = FALSE) gets the analogous fix (a zero weighted score maps to a zero unweighted score instead of 0/0). Outputs for all other interval ranges are unchanged.

Tests were written first and failed on the unfixed code:

FAILURE: 'test-metrics-quantile.R:627:3'
Expected `wis(5, predicted, quantile_level)` to equal 0.4.
  `actual`: NaN
`expected`: 0.4

FAILURE: 'test-metrics-quantile.R:746:3'
Expected `interval_coverage(observed, predicted, quantile_level, interval_range = 30)`
not to throw any conditions.
Actually got a <rlang_error> with message:
  ! To compute the interval coverage for an interval range of "30%",
    the 0.35 and 0.65 quantiles are required

After the fix, the full test suite passes (0 failed, 920 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 or user-facing documentation changes needed; only code comments were added).
• 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() on the changed files; no new lints were introduced.
• 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 (68aeedb).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1217   +/-   ##
=======================================
  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):

Verification

• Fail-before/pass-after confirmed: with R/ reverted to main (tests kept), exactly the 4 new tests fail (11 failures + 1 error) for the stated reasons — interval_coverage() aborts on seq() levels, and wis()/interval_score()/quantile_score(weigh = FALSE) return NaN for levels 0/1. On the branch everything passes.
• Full suite on the branch: 0 failed, 920 passed (matches the PR claim).
• Independent edge checks: weighted WIS = mean quantile score identity holds to 3e-15 over 500 random forecasts including levels 0/1 (also with the observation exactly on a bound); Inf locations agree between unweighted wis() and quantile_score(); new interval_score() is numerically identical to the old formula for ranges 0/10/50/90 (weighted and unweighted, incl. vectorised mixed ranges); NA propagation unchanged; interval_coverage() returns correct values (not merely no error) for seq() levels and still errors on genuinely missing quantiles. Brief decisions (10-dp rounding, option (a), sibling quantile_score fix, NEWS entry) are all respected; scope, PR title format, and NEWS entry are fine.

Non-blocking observations

1. Low: tolerance mismatch between the fixed presence check (quantile levels rounded to 10 dp) and the downstream row filter (interval ranges rounded to 10 dp; quantile error is amplified x200 in the range). A quantile level off by ~1e-11 now passes the check but silently returns NA instead of the previous loud error. Practically unreachable via the standard pipeline (as_forecast_quantile() pre-rounds levels to 9 dp), so noting only.
2. Informational: with weigh = FALSE and levels 0/1 present, wis() and quantile_score() are now both finite but differ by design (the 100% interval contributes its dispersion to wis() but 0 to the quantile scores at levels 0/1). This follows directly from the approved 0 -> 0 mapping; the documented identity is the weighted default, which the PR restores exactly. A short docs sentence could make this explicit.

Verdict: approve.