fix: composite key collision in multi-column joins + _accel.c v3

CLAUDE.md

@@ -67,7 +67,7 @@ The library has one core abstraction and a set of aggregation/partitioning opera
 
 **Supporting modules:**
 - `protocol.py` — Typing protocols for duck-typing compatibility (Series, DataFrame, Cursor)
-- `formatter.py` — `ObjectFormatter` for custom dtype parsing (e.g., Decimal → float); auto-converts object arrays of strings to `StringDType(na_object=None)` for nullable string support
+- `formatter.py` — `ObjectFormatter` for custom dtype parsing (e.g., Decimal → float); string conversion from object to `StringDType(na_object=None)` is opt-in via `parse_object_dtypes_inplace()`
 - `csvreader.py` — CSV reader with type inference; string columns produce `StringDType(na_object=None)` for nullable string support
 
 ## Testing

docs/changelog.md

@@ -1,5 +1,17 @@
 # Version History
 
+## 2.2.2
+
+* **Fix**: Multi-column join composite key collision when left and right sides
+  have different numbers of unique values. `_build_composite_key` now uses
+  shared cardinalities (max over both sides) so positional encoding is
+  consistent between left and right keys.
+* **Fix**: Constructor no longer auto-converts object arrays of strings to
+  `StringDType`. This preserves the caller's dtype and prevents `union()`
+  mismatches when one Tafra comes from an external source with object dtype.
+  Use `parse_object_dtypes_inplace()` for explicit conversion.
+* **Perf**: `_accel.c` v3 — combine bitmask hashing with slot-caching joins.
+
 ## 2.2.1
 
 * **Fix**: Join and Union dtype validation now compares `_dtypes` metadata

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "tafra"
-version = "2.2.1"
+version = "2.2.2"
 description = "Tafra: essence of a dataframe"
 readme = "README.md"
 license = "MIT"

recipe/meta.yaml

@@ -1,5 +1,5 @@
 {% set name = "tafra" %}
-{% set version = "2.2.1" %}
+{% set version = "2.2.2" %}
 
 
 package: