gg_varpro(): intermittent failure on survival fits — varPro importance() returns 0 rows

[ehrlinger]

Symptom

gg_varpro() on a survival varPro::varpro() fit can fail with:

Error: arguments imply differing number of rows: <p>, 0

(e.g. differing number of rows: 12, 0). This is real and not rare on real survival data, even though it does not reproduce on every fit (e.g. pbc across ntree = 50..200 happened to land on the working path during triage). gg_varpro() survival support otherwise works (the varpro vignette renders gg_varpro(v_pbc)), so this is an intermittent failure, not a categorical limitation.

Likely root cause

The cbind/data.frame in the importance reshape (.build_varpro_imp_dfs() → it combines the p variable names with the rows returned by varPro::importance()) gets a 0-row importance table for the failing survival fits. So varPro::importance() is returning an empty result for those survival fits — likely a varPro-side condition (no rules/variables survive the importance pre-filter, or a survival-specific code path).

Tasks (later / dev line)

1. Find the varPro condition. Build a reproducer (a survival fit where varPro::importance(fit) returns 0 rows), isolate what triggers it, and file upstream with varPro (Lu/Ishwaran) if it's a varPro bug.
2. Protect against it in ggRandomForests. Detect the empty-importance case in .build_varpro_imp_dfs() / gg_varpro() and fail gracefully with a clear, specific message (or a sensible fallback) — scoped to the 0-row case, NOT a blanket survival family block. A family block was tried in fix(gg_varpro): clear error on survival fits instead of cryptic cbind #116 and was wrong: it breaks the working survival path and the vignette render (fix(gg_varpro): clear error on survival fits instead of cryptic cbind #116 closed).

Context

• Surfaced while building the ggRandomForests 3.1.0 release-announcement figure (gg_vimp vs gg_varpro side-by-side).
• fix(gg_varpro): clear error on survival fits instead of cryptic cbind #116 (closed) attempted a blanket survival block — wrong approach; the fix here must preserve the working survival path.
• Targets the dev / v4.0.0 line.