…ze (#110)

The regression and survival partial-dependence surfaces were interactive
plotly widgets; self-contained quarto inlined plotly.js (~3.5 MB) into
each vignette HTML, and figures rendered at retina 2x. Installed size was
17.1 MB (doc 16.3 MB), well over CRAN's 5 MB guideline.

Replace both surfaces with static ggplot2 heat maps, set fig-format png /
fig-dpi 96 in all four vignettes, and drop the now-unused plotly Suggests.
Installed size drops to ~5.5 MB (doc 4.7 MB); source tarball 9.0 -> 3.7 MB.

R CMD check --as-cran (with manual, ggraph present): pending confirmation.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…113)

win-builder R-oldrelease flagged the plot.gg_variable example at 10.33s
elapsed (just over CRAN's 10s; under 10s on release/devel). Wrap the
loess-heavy regression panel plot and the full survival section (veteran
forest + multi-time variable/panel plots) in \donttest so they are
excluded from the timed example run; the fast classification + basic
regression plots still run. No behaviour change.

R CMD check --as-cran: examples [14s] OK, examples --run-donttest [28s]
OK, Status: OK (0/0/0).

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* perf(check): cut CRAN overall check time below 10 min

CRAN flagged the 3.1.0 submission's overall check time (13 min > 10 min),
driven by the vignette rebuild (331s) and tests (209s). Reduce both per
Uwe Ligges' suggested levers (toy data / fewer iterations / precomputed
results), with no change to test coverage or vignette content.

Vignettes:
- regression: Boston forest ntree 200, PD-surface grid 25 -> 10
- survival: impute ntree 100, forest ntree 150, PD-surface grid 25 -> 8
- varpro: the three gg_partial_varpro() calls (11-17s each) and the Boston
  beta.varpro() fit (~3s) -- the bulk of that vignette -- are precomputed
  offline by precompute_varpro.R and loaded from varpro_precomputed.rds
  (167 KB, xz), with an automatic live-computation fallback if absent.

Tests:
- test_gg_udependent memoised varPro::get.beta.entropy() (~1.5s, a pure
  function of the fit) per argument signature instead of recomputing it
  once per test (this file was ~24s of the suite, now ~9s).

Verified: R CMD check --as-cran with manual is OK (0/0/0); local vignette
rebuild 33s and tests 28s (were 331s/209s on CRAN's r-devel-windows).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* review: address Copilot feedback on the check-time PR

- varpro.qmd: load the precompute via tryCatch(readRDS) so a missing OR
  unreadable .rds falls back to live computation instead of erroring.
- precompute_varpro.R: mirror the vignette's requireNamespace/pkgload
  fallback instead of bare library(ggRandomForests), so the script runs in
  a fresh clone before the package is installed.
- test_gg_udependent.R: make make_ggu() warning suppression opt-in
  (.quiet = FALSE by default); pass .quiet = TRUE only to the empty-graph
  (threshold = 999) cases that legitimately warn, so an unexpected warning
  on any other call still fails the test.

Verified: test_gg_udependent 19 tests, 0 fail / 0 warn; varpro vignette
renders in 20s with 0 errors (precompute still loaded).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* review(#114): search both paths for varpro_precomputed.rds

The precomputed-load chunk read only the cwd-relative
'varpro_precomputed.rds'; depending on how Quarto sets the working
directory during R CMD check this could miss the file and silently fall
back to (slower) live computation. Search both the vignette-dir and
package-root locations before the live fallback. Addresses Copilot review.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

v3.1.0 accepted to CRAN (2026-06-11). Bump main to the post-release
.9000 dev version (DESCRIPTION + NEWS, dual update so the news-version
test sees the DESCRIPTION version), and record the release submission in
CRAN-SUBMISSION (SHA a7d8052).

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…p + vignette precompute) (#119)

* fix(cran): skip_on_cran the varPro tests to avoid upstream UBSAN (v3.1.1)

CRAN's gcc-UBSAN additional check flagged 3.1.0 with a 0-length array
access in randomForestSRC's compiled rfsrcGrow (entry.c:184), reached
when varPro::varpro()/beta.varpro() grow a forest in our tests.
ggRandomForests is pure R (NeedsCompilation: no) — the overflow is in a
dependency, surfaced by our tests.

Gate every varPro forest-growing test fixture/builder with
testthat::skip_on_cran() so CRAN's machines (incl. gcc-UBSAN) never run
them. No package code changed; the tests still run on CI and locally.
Bump to 3.1.1; NEWS + cran-comments explain the fix. Upstream reported.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(cran-comments): note the benign 'days since last update' NOTE

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(cran): precompute all varPro vignette fits to avoid upstream UBSAN

The varpro vignette grew ~10 varPro forests live (varpro/uvarpro/isopro/
ivarpro/beta.varpro), each reaching randomForestSRC's rfsrcGrow rule-grow
path that trips the gcc-UBSAN "0-length array" report (entry.c:184, a
length-0 yvar.wt decremented to an out-of-bounds pointer). Cache every fit
in vignettes/varpro_precomputed.rds and load it with a live fallback, so
R CMD check performs no live varPro grow.

Strip the unused embedded forests ($rf, $isoforest, redundant ivarpro
attrs) before saving — validated that every gg_* wrapper call returns
output identical to the un-stripped object — keeping the file at 414 KB
(tarball 4.13 MB, under CRAN's 5 MB limit).

R CMD check --as-cran (with manual): 0 errors, 0 warnings, 1 NOTE
(days-since-update, expected). Overall check time 2:43.

dev/randomForestSRC-ubsan-report.md records the upstream root cause +
suggested patch (maintainers are aware and staging a fix).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* style(vignette): satisfy lintr (brace/semicolon) in fallback chunks

Lint CI flagged 16 brace_linter/semicolon_linter issues in the vignette
hardening:
- varpro.qmd: the load-with-fallback chunks braced the `if` branch but not
  `else` (`} else .vp$x`); brace both branches to match the existing
  precomputed chunks.
- precompute_varpro.R: rewrite the one-line .strip_* helpers (compound
  semicolons + inline braces) as multi-line.

Behaviour-preserving; lint_package() now returns 0. The shipped
varpro_precomputed.rds is unchanged.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs: address Copilot review — accurate CI claims, scoped strip comments

Copilot flagged that skip_on_cran() also skips under R CMD check in CI
(NOT_CRAN unset in the workflows), so the varPro tests run nowhere in CI:
- NEWS.md / cran-comments.md no longer claim the tests "run on CI"; they
  state the tests run locally (devtools::test()) and are skipped under
  R CMD check, including the CI check jobs. (Restoring CI coverage via
  NOT_CRAN=true is deferred to 3.1.2, coupled with the issue #118 fix so
  the intermittent survival gg_varpro test doesn't flake CI.)
- precompute_varpro.R: scope the "forests unused" comments — the survival
  C-path gg_partial_varpro() and gg_isopro(newdata=) DO use $rf/$isoforest,
  but the vignette never invokes those on a stripped object (pd_pbc cached,
  gg_isopro training-path only).

No behaviour change; lint_package() = 0.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* ci: set NOT_CRAN=true so varPro tests run in CI (Copilot review)

Copilot noted skip_on_cran() also skips the varPro tests under R CMD check
in CI (NOT_CRAN unset), so they ran nowhere. Set NOT_CRAN=true in all
check/coverage workflows to restore that coverage. Safe: CI is not a
sanitizer build, so the upstream randomForestSRC UBSAN path is harmless
here; only CRAN's own check machines (NOT_CRAN unset) skip them, which is
what avoids the gcc-UBSAN additional issue. The CI-run varPro tests cover
regression + classification only (no survival), so issue #118 cannot flake
CI. Verified locally under NOT_CRAN=true: 0 failures across the varPro
suite.

NEWS.md / cran-comments.md restore the accurate "runs in CI and locally;
skipped only on CRAN" claim.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…BSAN trigger (#122)

* fix(cran): v3.1.2 — skip_on_cran the isopro fixture (clear gcc-UBSAN)

v3.1.1 guarded the varPro forest-growing test fixtures with skip_on_cran()
but missed make_iso_fit() in tests/testthat/test_gg_isopro.R. It was gated
only by skip_if_not_installed("varPro"), and varPro IS installed on CRAN's
check machines — so the ~23 isopro tests still ran there.

varPro::isopro() grows an *unsupervised* isolation forest (do.call("rfsrc",
...) with the unsupv family → yvar.wt length 0). That is the exact path that
trips randomForestSRC's gcc-UBSAN report at entry.c:184 (unconditional
RF_yWeight-- on a 0-length weight vector → out-of-bounds pointer, UB). It was
the one unsupervised grow v3.1.1 left unguarded, so the additional check
re-flagged the issue against 3.1.1.

Fix: add testthat::skip_on_cran() to make_iso_fit(), matching the other
varPro fixtures. ggRandomForests is pure R (NeedsCompilation: no); package
code is unchanged. The isopro tests still run in CI and locally
(NOT_CRAN=true).

Verified: R CMD check --as-cran (with manual) → 0 errors, 0 warnings, 1 NOTE
(days-since-update), 2.6 min. test-all.Rout shows all 23 isopro tests in the
"On CRAN" skip list, so the unsupervised grow no longer executes under the
check. Bumps DESCRIPTION + NEWS to 3.1.2; cran-comments updated.

Upstream fix is randomForestSRC commit 92ec283 (not yet on CRAN).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(cran): narrow the UBSAN skip to only isopro(method="unsupv")

Empirically pinned the gcc-UBSAN trigger with -fsanitize=undefined on CRAN's
randomForestSRC 3.6.2: of every varPro/rfsrc grow in the suite, ONLY
varPro::isopro(method="unsupv") fires entry.c:184. It is the sole grow with a
length-0 yvar.wt (unsupervised family). Everything else is supervised or
synthetic-supervised and UBSAN-clean:
  - varpro / ivarpro / beta.varpro: real Y -> non-empty yvar.wt
  - isopro(method="rnd"): synthetic supervised forest
  - uvarpro: grows yxyz123 ~ . with an rnorm response (synthetic supervised)

So:
  - make_iso_fit() now skip_on_cran() ONLY when method == "unsupv"
    (the rnd isopro tests run on CRAN again).
  - Reverted v3.1.1's blanket skip_on_cran() on the varpro/uvarpro/ivarpro/
    beta fixtures (helper-varpro-fixtures.R, test_gg_varpro.R,
    test_gg_udependent.R) -> ~126 varPro tests run on CRAN again.

Verified:
  - Full suite under CRAN conditions (NOT_CRAN unset) against a GCC-equivalent
    UBSAN rfsrc (-fsanitize=undefined -fno-sanitize=float-cast-overflow):
    0 runtime errors, FAIL 0 / PASS 1093 / SKIP 7. Only 2 tests skip On CRAN
    (isopro-unsupv + the pre-existing slow ivarpro test).
  - R CMD check --as-cran (with manual): 0 errors, 0 warnings, 1 NOTE
    (days-since-update); overall 2.68 min, well under the 10-min budget.

Note: a local clang UBSAN build also surfaces partial.c:92 (a (uint)NaN
float-cast in test_gg_partial_varpro's categorical survival partial
dependence). CRAN does not see it: GCC's -fsanitize=undefined excludes
float-cast-overflow (clang's includes it). Confirmed by rebuilding with that
check disabled. Separate latent upstream issue, not a CRAN blocker.

NEWS + cran-comments updated to describe the narrowed fix.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* test(isopro): namespace-qualify skip_* in make_iso_fit

Use testthat::skip_on_cran()/skip_if_not_installed() so the helper does not
depend on testthat being attached. Addresses a Copilot review note on #122;
qualifies both calls for internal consistency.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(cran-comments): correct NOTE disposition + narrowed-skip wording

R CMD check --as-cran returns 1 NOTE (days since last update), not 0; only the
single unsupervised isopro test skips on CRAN now (all other varPro tests run).
Genericise the day count so the submitted comment is not stale.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

submit_cran() recorded the submitted commit (SHA 83d8021, the merged #122)
and timestamp 2026-06-13. ggRandomForests 3.1.2 submitted to CRAN to clear the
gcc-UBSAN additional issue.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* fix(gg_partial_varpro): drive RMST(tau) partial computation, not just label (v3.2.0)

varPro::partialpro() has no time argument, so its default survival learner
returns ensemble mortality at every horizon. gg_partial_varpro(scale="rmst",
time=tau) previously only relabeled the y-axis, so multi-horizon RMST plots
differed by Monte-Carlo noise rather than tau.

scale="rmst" now passes partialpro() an RMST(tau) learner that integrates the
survival curve (integral_0^tau S(t) dt) from object$rf, so the curve genuinely
depends on tau. This recomputes from object (a survival fit) with
part_dta=NULL; a precomputed part_dta can only be relabeled, and the function
now warns when you try. Also warns when tau exceeds the model's event-time
range (RMST truncated there) and when time is passed to a scale that ignores it.

New internal helpers .rmst_learner() and .rmst_from_survival(); RMST math unit
tests run on CRAN, the heavy end-to-end varpro fit is skip_on_cran() to protect
the check-time budget. Minor release 3.1.2 -> 3.2.0.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(gg_varpro): clear error on degenerate varPro importance table (#118)

gg_varpro() failed with the cryptic "arguments imply differing number of
rows: <p>, 0" when varPro::importance() returns a degenerate importance
table (0 rows, or p variables with no usable z column) -- observed
intermittently on survival fits where the release-rule step selects no
variables. .build_varpro_imp_dfs() now detects that and stops with a clear,
specific message suggesting a larger ntree. Scoped to the degenerate case;
well-formed fits (survival included) are unaffected -- not a blanket
survival-family block (cf. the reverted #116).

Cherry-picked from the dev-line fix in 21805e4 (the gg_varpro.R guard +
its CRAN-safe unit test only, not the bundled uvarpro feature work).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* build: require varPro (>= 3.1.0) in Imports

The RMST(tau) partial path passes partialpro() a `learner`, an argument
present since varPro 3.1.0 (the current CRAN version). Pin the floor so the
dependency is explicit.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(gg_partial_varpro): address Copilot review on RMST guardrails

Two fixes from the PR #127 review:
- The tau-range warning fired for tau < min(time.interest), but a small tau
  is valid: RMST integration assumes S(t)=1 on [0, times[1]), so it is not
  truncated. Only warn when tau exceeds the largest event time (the actual
  truncation case, since S(t) cannot be extrapolated past max(ti)); message
  corrected accordingly.
- Validate that 'time' is a single finite numeric. It now drives the RMST
  integration (and the surv/chf snap) as a scalar; a vector would silently
  recycle and return incorrect results.

Tests added for both.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* test(gg_partial_varpro): cover RMST/scale internals; fix C-path model on empty frame

Addresses the Codecov patch-coverage comment on PR #127 (was 84% patch /
gg_partial_varpro.R -4.16%). Adds CRAN-safe tests for the RMST-path internals
that were uncovered: .resolve_varpro_scale() auto-by-family branches,
.rmst_from_survival() bare-vector input, the .rmst_learner() OOB + newdata
branches (via a small rfsrc fit, no varpro grow), the plain partialpro(object)
mortality recompute (folded into the skip_on_cran e2e), and the C-path model
label. File coverage 91% -> 99%.

The C-path model test surfaced a real bug: .gg_partial_varpro_cpath() assigned
a scalar `model` column to pd$continuous/pd$categorical, which errors on a
0-row data.frame ("replacement has 1 row, data has 0 rows") when a variable
yields an all-continuous or all-categorical split. Guarded with nrow > 0.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* style(gg_partial_varpro): cut .validate_varpro_inputs cyclomatic complexity

The RMST scalar-time and survival-fit checks pushed .validate_varpro_inputs to
cyclomatic complexity 28, over the lintr cyclocomp_linter limit of 20 (CI lint
failure on PR #127). Extract two helpers -- .validate_partial_time() and
.validate_rmst_inputs() -- per the linter's own suggestion. Behavior and error
messages unchanged; package lints clean, tests green.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* feat(gg_partial_varpro): forward ... to partialpro() (variable selection)

The object-driven path called partialpro(object) / partialpro(object,
learner=...) with no xvar.names, so variable selection fell back entirely to
varPro::get.topvars(object) -- which returns few or no variables for some
fits, yielding empty continuous/categorical frames. Because the RMST path must
recompute from `object` (it cannot accept a precomputed part_dta carrying the
learner), callers had no way to specify variables the way an explicit
partialpro(xvar.names=...) call would.

Add `...` to gg_partial_varpro() (and the gg_partialpro() alias), forwarded to
partialpro() on the recompute path, so xvar.names / nvar / cut / nsmp etc. work
again. Warns when `...` is passed alongside a precomputed part_dta (ignored).
Matches the ...-forwarding convention of the other varPro wrappers
(gg_beta_varpro, gg_ivarpro).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(gg_partial_varpro): align RMST integration to the prediction's own time grid

The RMST learner integrated predict.rfsrc()$survival against the closure's
rf$time.interest, but predict.rfsrc() can return survival on a different column
grid (observed on a real survival fit: newdata predictions came back ~40x too
small, e.g. RMST(365) = 5-9 vs the OOB 109-365 for the same subjects). The
mismatch did not error -- R's negative indexing (surv[, -n_times]) silently
misaligned the integral -- so partialpro built RMST partial curves on wrong
values while still looking populated.

Use each prediction's own $time.interest for the integration, and assert in
.rmst_from_survival() that ncol(surv) == length(times) so any future grid
mismatch fails loud instead of returning a wrong number.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(cran): refresh cran-comments.md for v3.2.0 submission

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(gg_partialpro): stop terse @param tags from overriding the shared Rd

gg_partialpro shares @Rdname gg_partial_varpro and sorts after it in collation,
so its "Passed to gg_partial_varpro" @param tags overrode the rich descriptions
on the canonical gg_partial_varpro.Rd page -- every argument rendered as the
circular "Passed to gg_partial_varpro", burying the real time/scale/... docs.
Drop the duplicate @param block from the alias (it shares gg_partial_varpro's
formals, so those definitions now populate the shared Rd). Addresses the
Copilot review on PR #127.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(cran): note win-builder R-devel/R-release 0/0/0 in cran-comments

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(cran): win-builder R-oldrelease also 0/0/0 (all three legs clean)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(cran): mac-builder 0/0/0 — all external builders clean

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…rv defaults) (#129)

* docs: open v3.3.0; fold in the RMST "Reading an RMST curve" interpretation section

Starts the 3.3.0 minor (off main) and pulls in the RMST-interpretation @section
that was prepared on docs/3.2.1-rmst-interpretation, re-versioned 3.2.1 -> 3.3.0.
The standalone 3.2.1 doc release is superseded by this cycle (the upcoming
varPro classification probability-scale figure work lands here too).

Also gitignore brainstorm/vignette/tarball build artifacts that were previously
untracked.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(spec): v3.3.0 design — interpretable y-axis scales for varPro partial plots

Classification partial plots default to probability (odds/logodds options);
survival gains an S(t) learner (scale='surv', user-supplied tau) and no longer
silently defaults to ensemble mortality. Causal contrast hidden on bounded
scales. Approved design, pre-implementation.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(spec): survival default tau — median follow-up, not an error (3.3.0)

Revise 3e: scale='auto'+survival now returns survival probability S(tau) at a
data-driven default tau (median follow-up), units-safe by construction; surv and
rmst both default tau when time is omitted. Replaces the earlier 'auto errors'
design. Mortality stays an explicit opt-in.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(spec): require explicit causal-curve documentation (what it is + when to use)

§4: dedicated section explaining causal = varPro's baseline-subtracted
virtual-twins estimator (Ishwaran & Blackstone 2025), when to use it (effect vs
level; cross-check), the not-a-structural-claim caveat, and why it's hidden on
bounded scales. Add the virtual-twins reference. §3d links to it.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(plan): v3.3.0 varPro partial-plot scales — TDD implementation plan

10 tasks, TDD per step: scale vocab/resolution, prob/odds conversion (mean of
probabilities), causal blanking, S(t) learner, median-follow-up default tau,
routing/validation, target-class labels, plot causal warning, docs/NEWS, full
gate. Grounded in current internals.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* feat(gg_partial_varpro): probability/odds/surv scales + S(t) learner + default tau (3.3.0)

Classification partial plots default to probability (scale auto->prob; odds/
logodds options; mean-of-probabilities back-transform). Survival defaults to
S(tau) via a new partialpro learner (auto->surv); surv/rmst default tau to
median follow-up (units-safe). Causal contrast blanked + dropped/warned on
bounded scales. Target-class provenance + new y-labels. Reconciled C-path tests
to chf; full file green (122).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(gg_partial_varpro): scale @details, prob/surv/causal @Sections, NEWS, virtual-twins ref

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(3.3.0): gg_partialpro scale vocabulary + escape [0,1] in roxygen

- gg_partialpro() alias scale default now matches gg_partial_varpro's 8-value
  vocabulary (was the old 5-value vector, breaking match.arg when forwarded).
- \eqn{[0, 1]} in two @details/@section blocks: markdown mode had turned the
  bare [0, 1] into a \link, tripping the Rd cross-reference check.
R CMD check --as-cran now clean (only the days-since-update + local HTML-tidy
NOTEs remain, both environmental).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* style(test): split compound skip_*() semicolons (lint)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(gg_partial_varpro): scale-aware messages for surv/rmst (Copilot review)

.validate_rmst_inputs renamed .validate_surv_scale_inputs and its stop message
now names the actual scale (was hard-coded 'rmst', misleading for surv).
.warn_varpro_rmst's precomputed-part_dta and tau-out-of-range warnings are now
scale-generic (were RMST-specific, misleading for surv).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(vignette): update varpro partial-dependence sections for 3.3.0 scales

Classification section now describes the probability default (P(Y=target),
causal hidden, odds/logodds options); survival section rewritten from the
stale 'C-path / ensemble mortality' prose to the S(tau) survival-probability
default via the partialpro learner, the units-safe median-follow-up tau, and
mortality/rmst/chf as explicit opt-ins. Fixed the family-support table cell and
two stale mortality/C-path mentions. Re-ran precompute so the bundled .rds
caches pd_iris_multi (prob) and pd_pbc (surv, tau=median follow-up); vignette
still loads from cache (no live grow, check-time unchanged).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(gg_partial_varpro): default tau only when recomputing; fix 'one-time' vignette wording (Copilot)

- The median-follow-up default tau now resolves only when part_dta is NULL
  (recomputing from object). A precomputed part_dta is label-only, so the old
  code emitted a misleading 'default horizon' message and could call
  .default_surv_tau() on a non-survival rf. Extracted to .resolve_default_tau()
  (also keeps gg_partial_varpro under the cyclocomp limit). Regression test added.
- Vignette: the tau message fires on every omitted-time call, not once; dropped
  the inaccurate 'one-time' wording.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

… V3 CRAN line (#130)

* feat: port gg_beta_uvarpro + gg_sdependent to the V3 CRAN line (3.4.0)

Brings the two uvarpro visualization wrappers (varPro::get.beta.entropy bar
chart; varPro::sdependent signal-variable lollipop) from the dev_rhf line down
to a V3 3.4.0 minor, per the V3=varPro / V4=randomForestRHF boundary. New files
only; no new dependency (varPro already in Imports). Tests carry skip_on_cran
for the uvarpro-grow gcc-UBSAN issue. Version 3.3.0 -> 3.4.0.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(vignette): add gg_beta_uvarpro + gg_sdependent to the uvarpro section (3.4.0)

Two subsections after gg_udependent, run live on the cached u_boston fit
(consistent with how gg_udependent is shown): the entropy-ranking bar chart and
the signal-variable lollipop. No precompute change (both are post-fit processors
on the existing uvarpro fit, like get.beta.entropy/sdependent already used by
gg_udependent in the vignette).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* docs(voice): voice-polish pass on the 3.3.0/3.4.0 prose (3.4.0)

Applies the writing-voice harness (persona: biostatistician/data-scientist) to
the spots a documentation scan flagged -- mostly the new 3.3.0/3.4.0 prose:
- RMST @section: drop the 'two consequences worth stating' meta-framing and the
  em-dash clusters; 'Two things follow. First... Second...'.
- survival vignette: 'unbounded relative-risk score, not a probability, works
  well as a single scalar' sterile balance -> plain 'not a probability; a single
  risk score for ranking'; vary the boilerplate tricolon (was verbatim-identical
  to the regression vignette).
- regression vignette: 'stable, low-variance estimator' -> 'smooth out the noise'.
- varpro vignette: gg_sdependent section now opens from the reader ('you have a
  ranking; now you want the cut line'), not the function; gg_varpro reading guide
  primed before the plot.
- README: authorless data blurb gets 'you'; drop the duplicated ggplot/patchwork
  gloss + tricolon (the 'Why?' section already explains it).
Prose only -- no code, no behavior, no version change.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(voice): fix grammar/accuracy from voice pass (Copilot review)

- varpro vignette: 'a tight box is a variable...' -> 'a tight box means every
  tree agrees on that variable'.
- README: restore the ggplot/patchwork gloss on the plot() line -- 'building on
  with +' was inaccurate for patchwork composites.
- RMST @section: 'buy more' -> 'buy you more' (missing object); Rd regenerated.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…ntry (#132)

* docs: add short uvarpro vignette; trim varpro; fix stale VignetteIndexEntry

Add a focused vignette for the unsupervised varPro wrappers and lift the
three unsupervised sections out of the (now supervised-only) varPro
vignette so there is one source of truth. Also fix a stale CRAN listing.

* New vignettes/uvarpro.qmd: walks one mtcars uvarpro() fit through
  gg_udependent() (structure), gg_beta_uvarpro() (ranking), and
  gg_sdependent() (the signal/noise cut), using the shared beta_fit
  matrix. Renders clean (3 figures incl. the ggraph network; citation
  resolves).
* varpro.qmd: reworked intro from "six wrappers" to "five supervised
  wrappers" with a pointer to the new vignette; replaced the three
  worked uvarpro subsections with a short pointer; fixed the broken
  cross-reference; removed the gg_udependent row from the family-support
  matrix.
* precompute_varpro.R: removed the now-dead u_boston fit, its list
  entry, and the stale comment.
* ggRandomForests.qmd: fixed \VignetteIndexEntry, which still carried the
  quarto template placeholder "Vignette's Title" (what CRAN listed) --
  now "Exploring Random Forests with ggRandomForests".
* .gitignore: ignore the uvarpro_* build artifacts (.Rbuildignore
  already covers them via wildcards).

No version bump: this folds into the accumulating 3.4.0 delta.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs: register uvarpro in pkgdown index; fix fragile §6 cross-refs

* _pkgdown.yml: add the new uvarpro vignette to both the top-level
  articles index (the missing entry that failed the pkgdown build:
  "1 vignette missing from index: uvarpro") and the navbar Articles menu.
* varpro.qmd: replace two hard-coded "§6" section references with
  relative wording ("the closing reference section", "the end of this
  vignette"). The vignette doesn't number sections, so §6 resolved to
  nothing (Copilot review). Also dropped a stale "deferred to v3.1.0".

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

#133)

* fix(gg_partial_rfsrc): impose factor levels by integer code, not label

partial.rfsrc() imposes a factor level by its internal integer code --
it runs as.numeric(partial.values) internally. gg_partial_rfsrc() was
passing the level *labels*, so character labels ("No"/"Yes") coerced to
NA and numeric-looking labels ("4"/"6"/"8") to out-of-range codes, and
every factor level collapsed to a single partial-dependence value (a flat
categorical partial plot; "NAs introduced by coercion" warnings).

Confirmed against ground truth (impose level on all obs, predict, mean)
and plot.variable(partial=TRUE): passing integer codes reproduces both
exactly; passing labels collapses all levels.

Fix: make_eval_grid() passes seq-of-level codes for factors and carries
the level vector; partial_one_var() relabels get.partial.plot.data()'s
numeric x back to labels. Continuous and numeric low-cardinality
predictors are unchanged. The stale comment claiming labels were required
is corrected.

Adds tests/testthat/test_gg_partial_rfsrc.R: factor levels must not
collapse, ordering is recovered, and per-level means track ground truth.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(gg_partial_rfsrc): return categorical x as factor in model level order

Addresses Copilot review on #133. Relabeling get.partial.plot.data()'s
numeric x to a plain character vector let the downstream factor(.data$x)
in plot.gg_partial_rfsrc() re-sort levels alphabetically (e.g. lo/mid/hi
-> hi/lo/mid), losing the model's level order.

split_partial_result() now returns categorical x as a factor with levels
in first-appearance order. The rows arrive blocked by level in ascending
code order (the model's level order), so this preserves it; factor(x) in
the plot method then keeps those levels. Done after the mixed
continuous/categorical rbind (x still character there) to avoid coercing
the numeric continuous x. @return doc and a level-order test added.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* test: rename single-letter L to lev (object_name_linter)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

#134)

Release gate passed on merged main (incl. #132 docs + #133 factor-partial
fix): R CMD check --as-cran WITH manual = Status OK (0/0/0), tarball 4.2 MB,
total check 3m19s (vignette rebuild 36s), urlchecker 17/17, 0 revdeps.

* cran-comments.md: rewrite for v3.4.0 (was stale at v3.2.0) -- interpretable
  varPro partial scales, unsupervised wrappers, the gg_partial_rfsrc factor
  fix; test envs and UBSAN NOTE disposition refreshed.
* DESCRIPTION: Date -> 2026-07-01.
* README.md docs sweep: add a pointer to the new uvarpro vignette (the
  unsupervised trio was undiscoverable from the README); refresh the stale
  "Recent changes" highlights (were stuck at v2.5/v2.6) to the v3.x line.

Version stays 3.4.0 (DESCRIPTION == NEWS; carries the 3.3.0+3.4.0 delta).

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

R CMD check flagged a non-standard top-level file `Rplots.pdf`. It's a
default-device artefact created when a plot is drawn with no open device
(interactive smoke-tests, testthat runs). It's already gitignored but was
not build-ignored, so R CMD build copied it into the tarball.

Add an unanchored `Rplots\.pdf$` rule so the artefact is excluded from the
tarball at the root or under tests/ regardless of who left one behind.
Verified: with strays present at both locations, the built tarball
contains no Rplots.pdf.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…beta_varpro, gg_ivarpro) (#136)

* fix(gg_isopro): add default S3 method, remove dead inner guard

gg_isopro had no default method: the friendly "expects an isopro object"
stop() lived inside gg_isopro.isopro, which is only reachable when the
object already inherits 'isopro', so it was dead code. A wrong-class input
got R's stock "no applicable method for 'gg_isopro'" error instead.

Add gg_isopro.default emitting the clear class error (matching the
gg_beta_uvarpro / gg_sdependent pattern) and drop the unreachable inner
check. Adds a dispatch-error test (needs no varPro fit, runs on CRAN).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(gg_beta_varpro,gg_ivarpro): add default S3 methods, remove dead guards

Same fix as gg_isopro, extended to its two varPro-family siblings so the
whole family is consistent with gg_beta_uvarpro / gg_sdependent. Each had an
unreachable `if (!inherits(object, "varpro")) stop(...)` inside its .varpro
method; a wrong-class input got R's generic "no applicable method" error.

Add gg_beta_varpro.default and gg_ivarpro.default emitting the clear class
error, drop the dead inner guards, and add fit-free dispatch-error tests
(run on CRAN). The 6 classic wrappers (gg_error/gg_vimp/gg_variable/
gg_rfsrc/gg_roc/gg_brier) still lack .default but have no dead guards and
are deferred to a post-release consistency pass.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* refactor(gg_beta_uvarpro): drop dead inner class guard

gg_beta_uvarpro already has a .default method (added with the wrapper), but
its .uvarpro method still carried the unreachable `if (!inherits(object,
"uvarpro")) stop(...)` guard. Remove it so the whole varPro family is
consistent (no dead guards, class errors handled by .default). Behavior is
unchanged — wrong-class input still errors via gg_beta_uvarpro.default.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

A local CLAUDE.md (repo-specific Claude instructions, references the
Obsidian codemap) sits at the repo root. It's not tracked, but nothing
kept R CMD build from sweeping it into the tarball -> a "non-standard file
at top level" NOTE, same class as the stray Rplots.pdf handled in #135.

Add `^CLAUDE\.md$` to .Rbuildignore (never ship it) and `CLAUDE.md` to
.gitignore (never commit it). Verified: with CLAUDE.md present in the
source tree, the built tarball contains no CLAUDE.md.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…wn) (#138)

The mac.r-project.org macOS builder is returning HTTP 502 at submission
time, so no mac-builder result exists. Replace the (now-untrue) "mac-builder
Status: OK" line with the honest statement: macOS is covered by the local
aarch64-apple-darwin23 --as-cran check and the macos-latest GitHub Actions
job. Also pin the win-builder R versions to the ones actually run
(R-devel r90199 / release 4.6.1 / oldrel 4.5.3), all Status: OK.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…3.4.1) (#139)

* fix: add default S3 methods to the classic rfsrc/randomForest wrappers (3.4.1)

Post-release consistency pass. gg_error(), gg_vimp(), gg_variable(),
gg_rfsrc(), and gg_brier() had no default S3 method, so a wrong-class input
got R's generic "no applicable method" error. Add default methods emitting a
clear "expected an 'rfsrc' or 'randomForest' object" message (naming the
class received), matching the varPro-family cleanup done in 3.4.0.

gg_roc() is intentionally left as-is: its default is aliased to gg_roc.rfsrc
(accepts rfsrc-shaped objects); changing that is a behavior change, not a
mechanical add.

Additive only — no change to valid-input behavior. New fit-free dispatch
test (runs on CRAN). Patch bump to 3.4.1 (DESCRIPTION + NEWS); will not be
submitted immediately (CRAN discourages back-to-back updates) — it starts the
post-3.4.0 dev line on main.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* chore(cran): record 3.4.0 submission (SHA 16c8b4d)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

Forward-merge the V3 CRAN line into the V4 (randomForestRHF) dev line so
dev_rhf carries main's accumulated fixes: interpretable varPro partial
scales (3.3.0), the gg_partial_rfsrc factor-code fix, the full dispatch
cleanup (varPro family + classic wrappers .default), the uvarpro vignette +
varpro.qmd trim, and the Rplots/CLAUDE.md build-ignore hygiene.

Conflict resolutions:
- DESCRIPTION: keep dev_rhf Version 4.0.0.9000; Date -> 2026-07-02. Both
  varPro (>= 3.1.0) and randomForestRHF (Suggests) retained.
- NEWS.md: dev_rhf's v4.0.0 (development) heading on top, then main's
  v3.4.1 / v3.4.0 / v3.3.0 sections; single Version: line (4.0.0.9000).
- NAMESPACE / man/: regenerated via document() -> union of main's .default
  methods (12 dispatch entries) and dev_rhf's RHF exports (gg_rhf/gg_auct).
- R/gg_partial_varpro.R, R/gg_beta_uvarpro.R: took main (0 RHF refs; 3.3.0
  scale work + reviewed uvarpro).
- R/print_helpers.R: kept dev_rhf's RHF provenance branch.
- test_gg_partial_varpro.R: main (3.3.0 scale tests).
- test_gg_vimp.R: kept dev_rhf's deterministic version (importance[1] <- -1).
- vignettes: took main's trimmed varpro.qmd + new uvarpro.qmd + precomputed
  rds (dropped u_boston); kept dev_rhf's MASS guard in precompute_varpro.R.
- .Rbuildignore / .gitignore / cran-comments.md / CRAN-SUBMISSION /
  survival.qmd: took main (functional superset / current truth).

Verified: package loads; test_default_dispatch 6/0, test_gg_vimp 63/0,
test_gg_partial_varpro 108/0.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>