v1.0.0 — gamstransfer backend

Major change — gamstransfer backend

• The package is now backed by gamstransfer,
  the GDX I/O package maintained by GAMS Software GmbH. No compiled code
  ships in this package any more (src/ removed).
• The high-level API is preserved: gdx(), [.gdx, extract(),
  all_items(), batch_extract(), write.gdx(), write2.gdx(). Existing
  user code calling these should not need changes.
• write.gdx() no longer shells out to a GAMS process; the removeLST,
  usetempdir and digits arguments are accepted for compatibility but
  ignored. compress = TRUE still produces a compressed GDX.
• write2.gdx() is now a thin alias of write.gdx() (both use the same
  fast path).
• igdx() and gams() are kept as thin compatibility shims (PATH-based
  GAMS discovery + system2 runner). The lower-level gdxrrw-shaped
  exports rgdx, wgdx, gdxInfo, rgdx.param, rgdx.set and
  rgdx.scalar have been removed.
• Behavior change: extracting a sparse parameter no longer densifies the
  result with explicit zeros. Only entries actually stored in the GDX are
  returned. (Variables and equations are unaffected.)

New options

• na = c("drop", "keep", "error") on write.gdx() / write2.gdx() —
  controls how NA / NaN values in parameter value columns are handled.
  Defaults to "drop" (legacy v0.7 semantic: silently discarded, GAMS
  reads 0 for those keys). "keep" preserves them as GAMS NA / undef;
  "error" stops with an informative message.
• dup = c("first", "last", "error") on write.gdx() / write2.gdx() —
  controls duplicate-key collapsing. Defaults to "first" (legacy
  write2.gdx / wgdx semantic, which auto-selected in 95% of v0.7
  calls). "last" matches the legacy write.gdx GAMS-process path;
  "error" stops when duplicates are present. A warning fires for both
  first and last so dropped rows are always audible.

Bug fixes

• write.gdx() drops default-valued records for free variables
  (level=0, lower=-Inf, upper=+Inf) to match legacy v0.7 semantic
  (subset(v, value!=0) for level, subset(v, !is.infinite(value))
  for bounds). Downstream tools (notably WITCH's witchtools time-period
  extension) rely on the absence of zero-level records to back-fill
  early years from the first non-default record.
• gdx() strips the position suffix gamstransfer adds to relaxed-string
  domain columns (witch13_3 → witch13). Callers look up the bare
  set name (df$witch13) and previously fell back silently when the
  suffix was present, which left WITCH parameters un-aggregated.
• write.gdx() warns when an index column is double-precision numeric
  with non-integer content. Catches a common scripting bug where a
  stray numeric column (e.g. low / high left over from a parquet
  read) silently becomes an extra UEL-string dimension. Integer-valued
  doubles (year stored as 1850.0) are still silent.
• write.gdx() collapses duplicate set rows the same way as parameter
  and variable keys.
• write.gdx() collapses duplicate index keys with dup = "first"
  (default) and warns with the dropped-row count.
• write.gdx() writes empty variables as a declaration of the right
  dimension instead of silently dropping them.
• write.gdx() handles duplicate index column names (e.g. (*, *, value)
  from a data.table) by make.unique-ing the records while preserving
  the original domain reference on the symbol.
• write.gdx() no longer leaks set column names as separate symbols
  in the output GDX (e.g. an explicit set r no longer produced a
  stray r parameter).
• write.gdx() pre-flight checks for NA values in index columns and
  stops with a clear error instead of letting gamstransfer's C++ layer
  abort the R session with SIGABRT.
• gdx() suppresses the gamstransfer acronym warning emitted at load
  time for symbols that use the GAMS Acronym facility.