missSBM v1.1.0

Major changes

• Breaking: estimateMissSBM()'s control must now be built with the new missSBM_param()
  (one named, defaulted argument per option) instead of a raw list(...), which
  now errors with a message pointing at the replacement.
• missSBM_fit now exposes split(), merge(), candidates_split() and candidates_merge()
  as instance methods, previously inlined in missSBM_collection's exploration logic; same
  search algorithm, now independently testable.
• new polish(control) (missSBM_fit and missSBM_collection): node-swap refinement after VEM
  convergence, fixing individually misclassified nodes that split()/merge() cannot reach.
• requesting more blocks than a network actually supports can make VEM collapse one or more
  classes; this used to be silent, and split/merge exploration's own repair of it could
  silently corrupt vBlocks's bookkeeping (duplicated/missing entries, non-smooth ICL/ELBO in
  plot()). Now fixed and made visible: missSBM_fit$repair(control) recovers a collapsed fit ;
  new occupiedBlocks/ degenerate fields expose any remaining collapse;
  bestModel skips degenerate models when possible and plot() marks them with a distinct point
  shape; estimateMissSBM()'s new stopOnDegenerate/maxConsecutiveDegenerate controls
  (default TRUE/2) stop forward exploration from growing further into a persistently unsupported range.
• new estimate_chain() (missSBM_collection) / warmChain control (default FALSE, opt-in):
  initializes each model by splitting the already-converged, smaller neighbor instead of an
  independent cold clustering, substantially reducing collapse in practice, at the cost of
  fitting vBlocks sequentially rather than in parallel.
• replace the NLopt/CCSAQ optimizer for the covariate connectivity parameters with a builtin
  Newton-Raphson solver (the M-step objective is concave, so it converges reliably in a handful
  of iterations); nloptr is no longer a dependency
• fix partlyObservedNetwork$imputation(): the fill value for missing dyads was biased low by
  including not-yet-imputed entries in its own computation. Known side effect:
  doubleStandardSampling_fit's initial psi bootstrap can degenerate when the fill value exactly
  matches the empirical observed edge rate (refined away by the VEM loop that follows)

Minor changes

• profiled estimateMissSBM() and fixed several hot-path inefficiencies: avoidable
  dense * sparse promotion overhead, a redundant recomputation of the dense imputed-network
  matrix, and an ifelse() evaluating both of its branches unconditionally. Roughly 3x faster
  on our benchmark, bit-identical results
• estimateMissSBM() now sorts/de-duplicates vBlocks if needed (with a warning), since
  exploration and chaining both assume it is strictly increasing.
• estimateMissSBM()'s control drops the exploration field: nobody used "forward"/
  "backward" alone at that level, only "both"/"none", so iterates now doubles as the
  on/off switch (0 disables exploration entirely). missSBM_collection$explore() no longer
  reads a stored exploration control field either; its direction argument (default "both")
  is now a plain per-call parameter, still available for advanced/internal use.
• cap the number of merge candidates tried during backward exploration
  (control$maxMergeCandidates, default 30) instead of always trying every pair
• fix a crash and a closed-form algebra bug in the "degree" sampling design; parameter recovery
  for this design can still be biased under heavy missingness (known limitation)
• fix a consistency bug in missSBM_fit$doVEM()'s step-back: only the SBM was restored, not the
  sampling model or the current imputation
• speed up getCovarArray() and kmeans_missSBM()'s seeding
• remove unused src/utils.h
• missSBM_fit's polish(), repair(), candidates_split() and candidates_merge() gain the
  same default control as doVEM(), so they can be called standalone without building a full
  control list first; missSBM_collection's calls (which always pass an explicit control list)
  are unaffected ; missSBM_collection's estimate()/polish()/explore() now share a control
  list stored at construction, no longer requiring control on every call.
• Factored missSBM_collection's explore_forward()/explore_backward() into a shared private
  explore_direction(), and missSBM_fit's candidates_split()/candidates_merge()'s trial-fit
  loop into a shared private trial_fit_candidates(); verified bit-identical output on the same
  fixed-seed scenario used throughout this changelog to diagnose VEM collapse