v2.0.0

2.0.0 is a breaking release. REF_AREA_CODE on the bulk SDMX path changes from the numeric IMF area code to ISO3 or a G-prefixed aggregate code, UNIT_CODE and the label columns follow the API codelists on both paths, and rows with a null OBS_VALUE are dropped. Historical row counts fall by roughly a third. REF_AREA_IMF_CODE carries the old numeric code for anyone migrating off it.

It also restores the eleven pre-April-2025 releases, which every version since the IMF put its download page behind bot management has been unable to fetch, and raises the Python floor to 3.12.

• weo.get_weo_versions() now reports every WEO release this package can fetch — the API's
  dataflow mapping plus the discontinued bulk SDMX archive (April 2019 through April 2025) —
  instead of only the two versions the API currently exposes. It is now exported from
  imf_reader.weo. get_weo_data(version=None) is unaffected: it still resolves "latest" against
  the API mapping alone.
• April 2021 and October 2023 are corrupt in the IMF's own published SDMX archive and cannot be
  fetched by any means. get_weo_versions() omits them; fetching one raises
  cache.BulkPayloadCorruptError with is_retryable=False and an explanation. See the README's
  "Coverage and known issues" section.
• The pre-April-2025 historical path (the bulk SDMX archive) is fetchable again. In v1.5.0 the IMF
  put its download page behind bot management and every one of those eleven releases raised
  ConnectionError; fetch_data can now reach all of them.
• The bulk SDMX path and the API path now return the same identifier vocabulary. REF_AREA_CODE
  is ISO3 (e.g. USA) or a G-prefixed aggregate code (e.g. G001) on both paths, where the
  bulk path previously returned the legacy numeric area code. UNIT_CODE, REF_AREA_LABEL,
  UNIT_LABEL, and CONCEPT_LABEL now follow the API's vocabulary and codelists on both paths
  too. This is a breaking change for the bulk SDMX path.
• Rows with a null OBS_VALUE are now dropped on both paths. Historical row counts fall by
  roughly a third, and about 7 series-level notes are dropped along with the rows that carried
  them.
• PPPGDP, PPPPC, PPPEX, and NGDPRPPPPC now have a null UNIT_CODE, where the legacy data
  carried T, F, and S. This is deliberate, not a regression: the IMF's own API publishes no
  unit at all for these four PPP / "international dollar" concepts, and CL_UNIT has no code for
  "international dollar" to translate to.
• LE, LP, and LUR carry a UNIT_CODE (PE, PT) on releases served from the bulk archive
  (April 2019 to April 2025) and a null one on releases served from the API (April 2025 onward).
  The API publishes no unit for these concepts either, and the bulk archive's unit depends on the
  area as well as the concept, so it cannot be carried forward to areas the API adds later. Code
  that reads UNIT_CODE for population, employment, or unemployment should not assume it is
  populated across versions; CONCEPT_CODE and CONCEPT_LABEL are stable for these series.
• A new REF_AREA_IMF_CODE column is added on both paths, carrying the legacy numeric IMF area
  code for each row (null for areas that never had one, e.g. LIE). It's a compatibility column
  for code migrating off the numeric area code, and is slated for removal in 3.0.
• imf-reader now requires Python 3.12 or later. Users on Python 3.10 or 3.11 keep resolving to
  the 1.5.x series.
• The default cache directory moves to a shared readerkit root: ~/.cache/readerkit/v1/imf-reader/<version>/
  on Linux, ~/Library/Caches/readerkit/v1/imf-reader/<version>/ on macOS, and
  %LOCALAPPDATA%\readerkit\Cache\v1\imf-reader\<version>\ on Windows. Users on Linux who want to
  reclaim disk space from the old directory can run rm -rf ~/.cache/imf_reader/ after upgrading.
• A new environment variable BBLOCKS_CACHE_DIR provides a family-wide fallback cache location,
  used when IMF_READER_CACHE_DIR is not set. Useful for pointing several packages at one shared
  cache root.
• HTTP requests now carry a default timeout and retry with backoff on transient failures and
  server errors.
• WEO bulk SDMX downloads now stream to disk and verify the transfer against the response's
  Content-Length header.
• The cache directory now uses readerkit's native file lock by default, instead of the
  SoftFileLock that imf-reader 1.5.0 always used. On a filesystem that rejects native locks,
  such as an NFS mount, readerkit raises CacheLockUnavailable and names the fix in the error.
  Set READERKIT_LOCK_STRATEGY=strict-soft to restore marker-file locking.
• A failed fetch can now raise ArtifactWriteError, CacheLockTimeout, CacheLockUnavailable,
  TruncatedDownloadError, SessionForkError, CacheDirectoryError, or RedirectPolicyError
  from the underlying caching library, in addition to the exceptions imf-reader already raised.
• weo.clear_cache(), weo.api.clear_cache(), and sdr.clear_cache() survive this release. v1.5.0
  announced their removal in 2.0; they still work and still emit a DeprecationWarning, and the
  removal moves to 3.0 alongside REF_AREA_IMF_CODE.
• The public imf_reader.cache API is otherwise unchanged.