1196: Fix as_forecast_quantile() from samples for asymmetric probs

[nikosbosse]

Description

This PR closes #1196.

as_forecast_quantile.forecast_sample() built the quantile_level column from a symmetrised vector unique(round(c(probs, 1 - probs), 10)) while computing the actual quantiles at the raw probs — a refactoring regression from 4d9ecf7 (#876), where the predecessor sample_to_quantile() used the same symmetrised vector for both. As a result, any probs not symmetric around 0.5 either silently paired quantile values with the wrong levels (via data.table length-1 recycling) or errored, and duplicated probs triggered a duplicate-forecast assertion.

The fix (per maintainer decision, Option B): compute quantiles at exactly the requested probs, using a single quantile_level <- unique(round(probs, digits = 10)) for both the column and the quantile() call — matching the documented behaviour of probs. For every symmetric prob set, including the default c(0.05, 0.25, 0.5, 0.75, 0.95), the output is byte-identical to before. The input check is also tightened to assert_numeric(probs, min.len = 1, lower = 0, upper = 1) for a clearer early error on out-of-range probs.

New tests failed as expected before the fix:

FAILURE: probs = 0.4 ->
  `actual`: 40.6 40.6      # two rows; level 0.6 silently got the 0.4 quantile
`expected`: 40.6

FAILURE: probs = c(0.1, 0.2) ->
  Error: Supplied 2 items for column 2 of group 1 which has 4 rows. ...

ERROR: probs = c(0.5, 0.5) ->
  Error: Assertion on 'data' failed: There are instances with more than one
  forecast for the same target. ...

All tests pass after the fix; the full suite passes on main + this change (0 failed, 916 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 changes; the fix makes the code match the existing documentation of probs).
• I have built the package locally and run rebuilt docs using roxygen2 (N/A — no roxygen comments changed, man/ files unaffected).
• My code follows the established coding standards and I have run lintr on the changed files (no new lints 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 (5cacd58).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1218      +/-   ##
==========================================
- Coverage   98.15%   98.15%   -0.01%     
==========================================
  Files          41       41              
  Lines        2225     2224       -1     
==========================================
- Hits         2184     2183       -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.

Checks performed (in an isolated detached worktree at e56c754):

• Fail-before/pass-after verified: with R/ reverted to origin/main (tests kept), all 4 new tests fail for the stated reasons (recycled 0.4 quantile onto level 0.6; data.table RHS-length error for c(0.1, 0.2); duplicate-forecast assertion for c(0.5, 0.5); quantile()'s internal error instead of the checkmate assertion for out-of-range probs). With the branch source restored, all pass.
• Full suite green: 872 passed, 0 failed, 18 skipped.
• Own edge-case probes (beyond the PR's tests): unsorted probs c(0.75, 0.2) across two forecast units pair levels and values correctly; boundary probs c(0, 1) return min/max; near-duplicate probs beyond the 1e-10 rounding precision dedupe to one row; default probs output is byte-identical to main, confirming no behaviour change for symmetric prob sets.
• Brief compliance: implements Option B (quantiles at exactly the requested probs, no symmetrization) and the approved lower = 0, upper = 1 bounds, as decided. Roxygen for probs already describes exactly the new behaviour, so no doc changes needed. Scope is minimal (source fix + tests + NEWS), NEWS entry accurate, PR title follows the NNNN: title convention, checklist truthful.

One minor observation (not blocking, pre-existing on main): probs = c(0.5, NA) passes assert_numeric() (bounds ignore NAs) and silently yields a row with quantile_level = NA / predicted = NA. Identical behaviour before and after this PR; any.missing = FALSE could be added in a follow-up if desired.

[nikosbosse]

Follow-up (Claude Fable, directed by @nikosbosse): addressed the review observation about missing values in probs.

probs = c(0.5, NA) previously passed the assert_numeric() check (checkmate bounds ignore NAs) and silently produced rows with quantile_level = NA / predicted = NA. Commit 5cacd58 adds any.missing = FALSE to the assertion in as_forecast_quantile.forecast_sample(), so NAs in probs now fail immediately with a clear assertion error, plus a regression test (expect_error on probs = c(0.5, NA)).

Verification: devtools::test(filter = "class-forecast-quantile") green (60 passed) and the full suite green (0 failed, 920 passed). Both changed files lint clean (source file linted with the package installed so object_usage_linter respects globalVariables).

[nikosbosse]

lgtm