v1.4.0

Changes:

• Made aligned full-file CICA Atlas subset requests explicitly return original
  files without applying dataset fixes by default. Deployments can set
  project:c3s-cica-atlas.apply_fixes_to_full_files = true to force those
  requests through normal processing and the configured fix backend. Whole-globe
  area parameters, and areas containing a dataset's complete spatial extent,
  are treated as no-ops when checking file alignment. Spatially disjoint areas
  now fail before subset processing starts.
• Added a memory-bounded post-processing step that merges consecutive subset
  and CMIP6-decadal concat batches up to the 2 GB output limit, returning one
  file when the complete result fits within that limit.
• Added independently configurable subset and decadal-concat batching memory
  aims, defaulting to 4GB. The planners now use the stricter of their
  timestep and estimated process-memory targets; concat accounts for every
  ensemble member.
• Moved adaptive time batching into the reusable rook.batch package and
  applied it to CMIP6-decadal realization concat so each time batch is written
  before the next one is assembled.
• Removed the legacy spec-file.txt alias; deployments should use the
  canonical linux-64.spec file directly.