diff --git a/NEWS.md b/NEWS.md
index 26f258e7..8f9248c2 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -3,6 +3,128 @@
 Development version, opened after 0.2.0.0 went to CRAN on 2026-08-03. Nothing
 released from this line yet.
 
+## Breaking changes
+
+* bgms requires R (>= 4.0). The package uses `tools::R_user_dir()` and builds
+  with C++20, both of which arrived in R 4.0; the declared floor was 3.5.
+
+* The `summary()` tables of a `bgmCompare` fit follow the same rownames
+  contract as `bgm()` tables, so `extract_rhat()` and `extract_ess()` return
+  named vectors; difference rows come contrast-major (all rows of one contrast
+  together) in both selection branches; and the slab block no longer carries
+  an `Rhat` column, which was structurally `NA` there. Code that indexed these
+  tables by position may need adjusting.
+
+## Bug fixes
+
+* `predict()` returned ordinal responses as the internal 0-based codes rather
+  than the categories of the data. `type = "response"` now maps predictions
+  back through the fit's recode map (including the Blume-Capel reference
+  shift), and probability columns are labelled with the original categories.
+
+* Predictions for rows of `newdata` carrying `NA` or an unseen category in a
+  conditioning variable were computed from corrupt internal codes. Those
+  entries now come back `NA`, with one warning naming the variables involved;
+  rows without such values are unchanged.
+
+* `predict()` errored on single-row `newdata` with `type = "response"` (a
+  dimnames mismatch specific to the one-row case).
+
+* `simulate()`'s ordinal full conditional subtracts the running maximum before
+  exponentiating, so extreme parameter draws no longer overflow, and an
+  interrupted parallel simulation errors instead of returning empty matrices.
+
+* `extract_centrality()` on mixed discrete-continuous fits summed the wrong
+  columns: pairwise draws are stored block-ordered (discrete, cross,
+  continuous pairs) and the strength sums indexed them as if they were
+  variable-ordered. Mixed-fit centrality values change; discrete-only and GGM
+  fits were indexed correctly.
+
+* `predict()`, `simulate()` and `extract_precision()` report the same
+  posterior-mean precision matrix, with the diagonal taken as the posterior
+  mean of the precision-diagonal draws, `E[K_jj]`. It was the reciprocal of
+  the posterior-mean residual variance, which is the harmonic mean of the same
+  draws. Diagonal entries grow slightly; `extract_partial_correlations()`
+  changes accordingly.
+
+* `prior_sensitivity_check()`'s importance weights carried only part of the
+  prior ratio, biasing the reweighted points between anchors (anchor points
+  are direct refits and were exact throughout). Two corrections: when the vary
+  mode also moves the precision-diagonal rate — the default
+  "slab-and-diagonal" resolution for GGM and mixed fits — the weights now
+  include the diagonal-prior ratio; and for mixed fits the diagonal statistic
+  in that ratio is read on the scale the sampler stores it (it entered with
+  the wrong sign and half the magnitude). Both affect curves that 0.2.0.0
+  computed for GGM fits and for mixed fits with continuous variables;
+  ordinal-only fits are unaffected. Validated against direct refits at a
+  doubled scale: the largest inclusion-probability error fell from 0.050 to
+  0.015 in the GGM check and from 0.051 to 0.006 in the mixed check, with most
+  edges inside twice the refit's own Monte Carlo error. The importance ESS is
+  computed from the corrected weights, so between-anchor points can report a
+  smaller ESS than before and be masked by `ess_floor`.
+
+* A saturated anchor — a reweighted inclusion probability at exactly 0 or 1 —
+  no longer takes over the pooled sensitivity curve: the inverse-variance
+  pooling weight uses an add-half smoothed probability. Pooled values in
+  non-saturated configurations move by less than one percent.
+
+* The stability interval starts from the chosen-scale verdict — the one the
+  table and `print()` report — rather than the pooled-curve verdict; the two
+  can disagree in borderline cases.
+
+* `prior_sensitivity_check()` input handling: single-indicator fits no longer
+  lose a dimension and error; slab families other than normal and Cauchy stop
+  with a clear message before any refit; a missing or non-positive chosen
+  scale is caught up front; and when a failed convergence gate is overridden
+  so the 1x anchor still contributes, `$grid` records the true gate result
+  together with a `forced` flag.
+
+* `calibration_check()` rejects `newdata` values outside the fit's support,
+  naming the variable and the offending values; both branches previously
+  failed silently (`NA` through the isotonic fit in the ordinal branch, silent
+  clamping in the Blume-Capel branch).
+
+* The Z-ratio truncation check runs on every table build; it was gated
+  together with the `alpha != 1` shape warning, so the default-shape cells it
+  was built for went unchecked. The shape warning keeps its gate.
+
+* Deployed correction-table slopes for the hierarchical and SBM graph priors
+  are computed after aggregating tied grid points: isotonic regression
+  produces runs of equal fitted density, and taking the slope over the tied
+  values biased it — a quarter of a nat on the measured cell.
+
+* Correction-table cache keys include the table build's own base seed (an
+  internal constant; the fit's `seed` does not enter the key), so tables built
+  under different base seeds are not confused; a failed cache write warns and
+  continues instead of erroring the fit; and table builds size their core
+  count independently of the chain sampler's configuration and announce it.
+
+* A fit with the Stochastic-Block edge prior and fewer than three continuous
+  variables warns that its Stochastic-Block updates run without the
+  normalizing-constant correction, since the slope curve is not resolvable for
+  a single pair, instead of building a table whose slope curve is empty.
+
+* Every chain the package launches resolves its acceptance target from one
+  helper: 0.44 for adaptive-Metropolis, 0.80 for NUTS, none for exact Gibbs.
+  The prior-only reference chains in `sample_ggm_prior()` and behind
+  `extract_prior_inclusion_probabilities()` previously fell back to the
+  compiled default of 0.80 — the NUTS target — and tuned their Metropolis
+  proposals differently from the fitted chains they mirror.
+
+* NUTS diagnostics tolerate non-finite energy values: E-BFMI is computed over
+  the finite draws and the warmup-incomplete flag no longer compares against
+  `NA`.
+
+* Adaptive-Metropolis diagnostics labelled the between-model move rate as the
+  acceptance rate. The acceptance probability the sampler records is now the
+  acceptance column; the move rate keeps its own column under its own name.
+
+* Named `newdata` columns that do not match the fit's variables stop with an
+  error (unnamed input keeps a warning); the legacy category-recode fallback
+  errors instead of silently shifting categories; `nsim` and `iter` are
+  validated at entry; and `sample_graph_prior()`'s joint-specification
+  metadata reports the effective Bernoulli prior used for conditioning.
+
 ## Diagnostics
 
 * `fit$nuts_diag$warmup_check` reports two new per-chain fields. `energy_tau` is
@@ -64,10 +186,10 @@ released from this line yet.
   ssgraph 1.16 were run on the vignette's own data, one at a time on one
   machine, each at its defaults and at its documentation's recommended
   settings. The vignette reports the headline numbers; the full protocol,
-  seeded runner scripts, and tables live in the bgms-docs repository beside
-  the site's comparison page, so the comparison can be re-baked when a
-  competitor's version moves. Neither competitor enters the package's
-  dependencies. The README summarizes the result in two sentences.
+  seeded runner scripts, and tables live with the site's comparison page, so
+  the comparison can be re-baked when a competitor's version moves. Neither
+  competitor enters the package's dependencies. The README summarizes the
+  result in two sentences.
 
 * `prior_sensitivity_check()`'s Details said the refit gate leans on per-chain
   verdict agreement and the indicator transition ESS. The gate reads the median
@@ -83,6 +205,24 @@ released from this line yet.
   states that the muted background holds robust and unlabelled scale-dependent
   edges alike, and that `print()` carries the counts and the names.
 
+* The ADHD dataset's DOI link renders correctly (the identifier was
+  double-prefixed) and the licence links are wrapped as URLs.
+
+* `calibration_check()`'s manual page describes the 95% PIT band as pointwise;
+  it said simultaneous, which the construction is not.
+
+## Other changes
+
+* The configure scripts stop at the first broken step: an unavailable `Rscript`
+  or an empty generated source list ends the build with a one-line message,
+  where they previously exited successfully and left `make` to fail later with
+  an unrelated error. `src/sources.mk` has a single writer.
+
+* Unreachable code was removed across `src/` and `R/`: a drifting ~190-line
+  duplicate of the bgmCompare gradient, three orphaned OMRF methods, unused
+  sampler plumbing, and small R-side leftovers. Sampler output is bit-identical
+  before and after the removals.
+
 # bgms 0.2.0.0
 
 This release rebuilds most of the package on top of 0.1.6.3. `bgm()` now fits
diff --git a/vignettes/fast-ggm.Rmd b/vignettes/fast-ggm.Rmd
index 86a72862..0efbc97d 100644
--- a/vignettes/fast-ggm.Rmd
+++ b/vignettes/fast-ggm.Rmd
@@ -355,8 +355,7 @@ data.
 How does the fast route compare with these two packages? A measured comparison
 is maintained on the package website's
 [comparison page](https://bayesian-graphical-modelling-lab.github.io/bgms/guide/fast-ggm.html),
-with the full protocol, seeded scripts, and complete tables
-[in its repository](https://github.com/Bayesian-Graphical-Modelling-Lab/bgms-docs/tree/main/benchmarks/ggm-comparison).
+together with the full protocol, seeded scripts, and complete tables.
 It lives there rather than here because its results depend on the current
 versions of the other packages, which change more quickly than `bgms` releases.
 
@@ -390,11 +389,15 @@ comparison section above.
 
 ```{r, echo = FALSE}
 data.frame(
-  field = c("CPU", "cores available", "threads used", "chains",
-            "operating system", "R", "bgms", "run"),
-  value = c(B$machine$cpu, B$machine$cores_total, B$machine$threads,
-            B$machine$chains, B$machine$os, B$machine$r_version,
-            B$machine$bgms_version, B$machine$date)
+  field = c(
+    "CPU", "cores available", "threads used", "chains",
+    "operating system", "R", "bgms", "run"
+  ),
+  value = c(
+    B$machine$cpu, B$machine$cores_total, B$machine$threads,
+    B$machine$chains, B$machine$os, B$machine$r_version,
+    B$machine$bgms_version, B$machine$date
+  )
 )
 ```