Ordered fix

[xrobin]

Summary

Fixes #63. Ordered predictors such as aSAH$wfns are no longer coerced to
numeric codes. Thresholds are the factor levels plus an Inf sentinel
("Inf" or "-Inf" depending on direction), and lookup is by label.
This is a breaking change for code that reads roc$thresholds or
passes numeric cut-offs into coords(), ci.thresholds(), ci.coords(),
or plot.roc(print.thres) on an ordered ROC. Sensitivity, specificity,
and AUC match the old numeric conversion; only the cut-off labels change.

# Before (pROC ≤ 1.19.1): named levels became 1:5, thresholds were midpoints
coords(roc(aSAH$outcome, aSAH$wfns), 2.5)
# After: pass the level
coords(roc(aSAH$outcome, aSAH$wfns, quiet = TRUE), "3")

Numeric predictors are unchanged (midpoints, coords(roc, 0.55), etc.).
Also fixes ci.coords(..., x = "best", best.policy = "omit"), which
returned the first column of the coords data.frame instead of dropping
the replicate. That bug was not ordered-specific.

Behavior

• An observation equal to the threshold is classified as positive
  (>= when direction = "<", <= when direction = ">").
• direction = "<": thresholds are c(levels, Inf); ">": c(-Inf, levels).
• Unused defined levels are kept as thresholds (they share SE/SP with the
  next occupied level toward the positive class).
• cases and controls must have identical levels(). Ordered
  cases/controls now build predictor/response in the same order
  (controls then cases), which fixes non-stratified bootstrap and
  Venkatraman on that interface.
• ci.coords() omits the threshold interval by default on ordered curves
  (labels are not numeric). ret = "threshold" is still accepted, returns
  NA, and warns.
• Exact level labels take precedence over keyword prefixes ("a" vs "all").
  If a string is both a keyword and a level, the keyword wins with a warning.
• Smoothing of ordered curves is binormal only.

Test plan

• roc(aSAH$outcome, aSAH$wfns) thresholds are ordered levels + Inf,
  not midpoints; AUC equals roc(..., as.numeric(wfns)).
• Named levels ("very low" … "very high") round-trip through
  coords(), ci.thresholds(), ci.coords(), plot(print.thres).
• Numeric x / print.thres / thresholds on an ordered ROC errors.
• Numeric ROC (ndka, s100b) still uses midpoints; coords(r, 0.55)
  and plot(r, print.thres = 0.55) unchanged.
• ci.coords(r_tie, "best", best.policy = "omit") on a curve with
  several Youden points drops those replicates (row of NAs) instead
  of returning the threshold column.
• devtools::test() / R CMD check
• Run reverse dependency checks and report summary/results.

Post-test communication

• Publish a blog post outlining user-facing ordered-threshold changes.
• Prepare and send an email to maintainers of reverse dependencies with migration guidance.

[codecov]

Codecov Report

❌ Patch coverage is 91.45299% with 30 lines in your changes missing coverage. Please review.
✅ Project coverage is 87.52%. Comparing base (d7b1962) to head (087dfaf).

Files with missing lines	Patch %	Lines
R/roc.utils.R	84.50%	20 Missing ⚠️
R/coords.R	97.25%	3 Missing ⚠️
R/ci.coords.R	90.00%	2 Missing ⚠️
R/ci.thresholds.R	87.50%	2 Missing ⚠️
R/obuchowski.R	50.00%	2 Missing ⚠️
R/roc.R	94.74%	1 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop     #142      +/-   ##
===========================================
+ Coverage    87.36%   87.52%   +0.16%     
===========================================
  Files           40       40              
  Lines         4146     4312     +166     
===========================================
+ Hits          3622     3774     +152     
- Misses         524      538      +14     

☔ 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.