@@ -23,10 +23,12 @@ jobs:
           python-version: ${{ matrix.python-version }}
       - name: Build wheel and source archive
         run: |
-          python -m pip install build==1.2.2
+          python -m pip install build==1.2.2 twine
           python -m build
       - name: Check archive contents
-        run: python maintenance/check-distribution-archives.py dist
+        run: |
+          python maintenance/check-distribution-archives.py dist
+          python -m twine check --strict dist/*
       - name: Install and exercise each distribution separately
         shell: bash
         env:
 
@@ -137,6 +137,9 @@ completed repository repair programme.
   alpha and are tracked as follow-up work below.
 - [ ] Review outstanding release metadata, archive size and service/citation links.
 - [x] Prepare the release notes for `0.12.0a1` in `doc/UserGuide/upgrade.rst`.
+  Release preparation now includes the completed Stage 9 previews, sorting and
+  retained uniqueness contracts. CI package validation checks the description
+  with strict Twine validation and reports actual archive sizes for review.
 - [ ] When authorised, tag and publish; record successful package/documentation
   jobs and verify published versions before closing release preparation.
 
 
@@ -29,7 +29,21 @@ The :doc:`datafile` and :doc:`image` guides describe the current interfaces.
 These native computations do not automatically apply Stoner exclusions. Use
 ``to_numpy()`` for a detached masked numerical export, ``Data.asarray()`` or
 ``ImageFile.imarray`` for read-only numerical snapshots, and ``export_storage()``
-for lossless state with typed metadata. Broader wrapper simplification is Stage 9.
+for lossless state with typed metadata.
+
+``Data.head(n=5)`` and ``Data.tail(n=5)`` return independent objects of the same
+class, preserving raw excluded measurements, masks, dtype, fill value, column
+headers/roles/IDs and typed metadata. Negative counts omit rows from the opposite
+end; zero returns an empty table. Standard previews copy only selected numerical
+rows, while custom clone/deepcopy hooks retain their established behaviour.
+
+``Data.sort`` continues to sort in place and return the same object. Compatible
+real numeric keys now delegate to pandas stable sorting, with exclusions moved
+by the same row permutation. Complex values, other float widths, non-native byte
+orders and custom comparators retain numerical paths. ``reverse=True`` reverses
+the complete ascending permutation, including ties and NaNs. ``Data.unique``
+retains NumPy's sorted masked results and optional index/inverse outputs because
+pandas distinct-value operations have different semantics.
 
 Direct native edits may change existing cells without changing dtype or geometry.
 Use Stoner methods for structural changes, and reacquire native references after
 
@@ -73,7 +73,10 @@ allocations versus masked exports and records regional selection costs and a
 runtime-source fingerprint. These checks supplement the scientific regression
 suite; they do not replace hosted platform or installed-distribution validation.
 
-`check-distribution-archives.py dist` inspects CI-built wheel/sdist contents.
+`check-distribution-archives.py dist` inspects CI-built wheel/sdist contents and
+reports compressed/uncompressed sizes and the largest retained assets. The package
+workflow also runs `twine check --strict` on both archives to validate their
+metadata and rendered package description before installed-distribution checks.
 `check-installed-package.py <checkout>` runs with each installed distribution from
 outside the checkout, checking import provenance, metadata, resources and sample
 data. The `Package validation` workflow provides those fresh environments; use its
 
@@ -40,10 +40,20 @@ def main():
         raise RuntimeError(f'Expected one wheel and one source archive, found {wheels}, {sources}')
     with zipfile.ZipFile(wheels[0]) as archive:
         check(archive.namelist(), wheel=True)
+        report_size(wheels[0], [(item.filename, item.file_size) for item in archive.infolist()])
     with tarfile.open(sources[0], 'r:gz') as archive:
         check(archive.getnames(), wheel=False)
+        report_size(sources[0], [(item.name, item.size) for item in archive.getmembers() if item.isfile()])
     print('Wheel and source archive content checks passed.')
 
 
+def report_size(path, entries):
+    """Report actual CI archive sizes and their largest retained source assets."""
+    print(f'{path.name}: {path.stat().st_size:,} compressed bytes; '
+          f'{sum(size for _, size in entries):,} uncompressed bytes; {len(entries)} entries')
+    for name, size in sorted(entries, key=lambda item: item[1], reverse=True)[:5]:
+        print(f'  {size:>12,} bytes  {name}')
+
+
 if __name__ == '__main__':
     main()