feat(dialect): follow the core's column-layer rebuild — column types, node formatters, and the version gates

.github/workflows/test.yml

@@ -18,8 +18,8 @@ permissions:
   contents: read
 
 jobs:
-  benchmark-python-314-postgres-18:
-    name: Run benchmarks with Python 3.14 / postgres 18
+  benchmark-python-315-postgres-18:
+    name: Run benchmarks with Python 3.15 / postgres 18
     runs-on: ubuntu-latest
     permissions:
       contents: read
@@ -56,10 +56,10 @@ jobs:
       - name: Check out the repository
         uses: actions/checkout@v6
 
-      - name: Set up Python 3.14
+      - name: Set up Python 3.15
         uses: actions/setup-python@v6
         with:
-          python-version: "3.14"
+          python-version: "3.15"
 
       - name: Install dependencies
         run: |
@@ -69,7 +69,7 @@ jobs:
       - name: Install activerecord and testsuite
         run: |
           pip install "rhosocial-activerecord @ git+https://github.com/rhosocial/python-activerecord.git@release/v1.0.0.dev30"
-          pip install "rhosocial-activerecord-testsuite @ git+https://github.com/rhosocial/python-activerecord-testsuite.git@feature/expression-schema-object-refactor"
+          pip install "rhosocial-activerecord-testsuite @ git+https://github.com/rhosocial/python-activerecord-testsuite.git@test/typed-column-contracts"
 
       - name: Create postgres scenario config
         run: |
@@ -181,9 +181,9 @@ jobs:
         if: always()
         run: |
           {
-            echo "## Python 3.14 / postgres 18 Benchmark Conclusion"
+            echo "## Python 3.15 / postgres 18 Benchmark Conclusion"
             echo
-            echo "- Python: 3.14"
+            echo "- Python: 3.15"
             echo "- postgres: 18"
             echo "- Scope: testsuite generic benchmarks + ActiveRecord bulk benchmarks + PostgreSQL backend direct benchmarks"
             echo "- Size: small"
@@ -202,9 +202,9 @@ jobs:
         if: always()
         run: |
           {
-            echo "## Python 3.14 / postgres 18 FastAPI Benchmark Conclusion"
+            echo "## Python 3.15 / postgres 18 FastAPI Benchmark Conclusion"
             echo
-            echo "- Python: 3.14"
+            echo "- Python: 3.15"
             echo "- postgres: 18"
             echo "- Scope: testsuite FastAPI database benchmarks"
             echo "- Size: small"
@@ -219,20 +219,24 @@ jobs:
             echo '```'
           } >> "$GITHUB_STEP_SUMMARY"
 
-  test-experimental-python-315:
-    # Temporarily disabled until Python 3.15 toolchain compatibility stabilizes.
-    if: ${{ false }}
-    name: Python ${{ matrix.python-version }} / postgres ${{ matrix.postgres-version }} (experimental)
+  test-python-315:
+    name: Python ${{ matrix.python-version }}${{ matrix.freethreaded && 't' || '' }} / postgres ${{ matrix.postgres-version }}
     runs-on: ubuntu-latest
     permissions:
       contents: read
-    continue-on-error: true
     strategy:
       fail-fast: false
       matrix:
         include:
-          - { python-version: '3.15.0-rc.2', postgres-version: '18', freethreaded: false }
-          - { python-version: '3.15.0-rc.2', postgres-version: '18', freethreaded: true }
+          # The free-threaded marker is in the job name even though there is
+          # currently no free-threaded leg, because two legs named alike is how
+          # a GIL pass and a free-threaded failure came to look like one job
+          # that both passed and failed.
+          #
+          # psycopg does not support free-threaded builds -- tracked as
+          # psycopg/psycopg#1095 -- so the GIL leg is the only one this backend
+          # has, on 3.15 as on every other interpreter.
+          - { python-version: '3.15', postgres-version: '18', freethreaded: false }
 
     services:
       postgres:
@@ -280,7 +284,7 @@ jobs:
       - name: Install activerecord and testsuite
         run: |
           pip install "rhosocial-activerecord @ git+https://github.com/rhosocial/python-activerecord.git@release/v1.0.0.dev30"
-          pip install "rhosocial-activerecord-testsuite @ git+https://github.com/rhosocial/python-activerecord-testsuite.git@feature/expression-schema-object-refactor"
+          pip install "rhosocial-activerecord-testsuite @ git+https://github.com/rhosocial/python-activerecord-testsuite.git@test/typed-column-contracts"
 
       - name: Create postgres scenario config
         run: |
@@ -392,7 +396,7 @@ jobs:
       - name: Install activerecord and testsuite
         run: |
           pip install "rhosocial-activerecord @ git+https://github.com/rhosocial/python-activerecord.git@release/v1.0.0.dev30"
-          pip install "rhosocial-activerecord-testsuite @ git+https://github.com/rhosocial/python-activerecord-testsuite.git@feature/expression-schema-object-refactor"
+          pip install "rhosocial-activerecord-testsuite @ git+https://github.com/rhosocial/python-activerecord-testsuite.git@test/typed-column-contracts"
 
       - name: Create postgres scenario config
         run: |
@@ -490,7 +494,7 @@ jobs:
       - name: Install activerecord and testsuite
         run: |
           pip install "rhosocial-activerecord @ git+https://github.com/rhosocial/python-activerecord.git@release/v1.0.0.dev30"
-          pip install "rhosocial-activerecord-testsuite @ git+https://github.com/rhosocial/python-activerecord-testsuite.git@feature/expression-schema-object-refactor"
+          pip install "rhosocial-activerecord-testsuite @ git+https://github.com/rhosocial/python-activerecord-testsuite.git@test/typed-column-contracts"
 
       - name: Create postgres scenario config
         run: |

changelog.d/62.added.md

@@ -0,0 +1 @@
+Add the PostgreSQL answer to the core column-type protocol and the four new pad/repeat/trim nodes. `PostgresColumnTypeMixin` and `POSTGRES_COLUMN_TYPES` carry the protocol's closed list of eighteen common Python types with every entry measured: `list`/`tuple`/`set`/`frozenset` answer `ArrayColumn` (native arrays, measured on 9.6/12/18 with the whole guarantee set green — `INT[]` DDL, a Python list bound directly and read back, `array_length`, 1-based `x[1]`, containment `x @> ARRAY[v]`, concatenation, `v = ANY(x)`, whole-column equality and `unnest`), `dict` answers `JSONColumn` while the DDL layer chooses `jsonb` through the `DataType` layer, `enum.Enum` answers `StringColumn`, and `float`/`decimal.Decimal` answer `NumericColumn` now that core removed `FloatColumn`/`DecimalColumn`/`RealColumn`/`DoubleColumn` and the numeric family is deliberately one class. `datetime.timedelta` is refused on purpose: PostgreSQL spells it `interval`, but no core column class names an interval's operations, and the previous `NumericColumn` answer offered `sqrt(interval '1 day')` for the engine to refuse at execution — a field declared `timedelta` now fails to resolve with a message naming the fix (`UseColumnType`). `PostgresDialect` composes the generic `TrimMixin`/`LpadMixin`/`RpadMixin`/`RepeatMixin` with no override, because PostgreSQL spells all four natively (`TRIM([BOTH|LEADING|TRAILING] [chars] FROM str)`, `LPAD`/`RPAD(str, len, padstr)`, `REPEAT(str, n)`). `PostgresUUIDMixin` and `PostgresUUIDAdapter` state the version-13-built-in-or-uuid-ossp answer and win over the core adapter's single fixed spelling, and `supports_uuid_ossp_extension` names the narrower question of whether uuid-ossp is installed.

changelog.d/62.changed.md

@@ -0,0 +1 @@
+Change what the dialect answers and where the protocol bases sit. `format_cast_expression` renders its target through `expr.target_type.to_sql()` instead of interpolating it, so `col.cast(DecimalType(dialect))` produces `"col"::NUMERIC` rather than `"col"::DecimalType()`, and the `type_values` and expression examples that showed a string target now show a type object. `JSONSupport` is no longer listed in `PostgresDialect`'s protocol block alongside its own derived `PostgresJSONBEnhancedSupport` — a base listed ahead of its own subclass is not a consistent MRO, and the pair had put `supports_json_type` in the MRO four times (two abstract protocol declarations, the core mixin default and the one implementation); the protocol still reaches the MRO, so `isinstance(dialect, JSONSupport)` is unchanged. `PostgresTableMixin`'s data-type validation accepts underscores in a type name, which no valid PostgreSQL type spelling was previously allowed to contain. `PostgresSchemaDiffer` records the consequence of comparing snapshots against an introspected array column, which always reports `dimensions=1` because PostgreSQL normalises every array declaration to a single dimension on storage.

changelog.d/62.fixed.md

@@ -0,0 +1 @@
+Fix the probes that promised a server answer the server does not give, and two mixins listed in an order that made their own implementation dead code. `supports_data_type_postgres_jsonpath`, `postgres_xid8`, the six `postgres_*multirange` probes and `postgres_macaddr8` were unconditional `True` on types that arrived in PostgreSQL 12, 13, 14 and 10 respectively — a gate that cannot say no tells the dialect it can render the type on any server, and the disagreement arrives at execution, with the DDL built and sent, and the error naming a table rather than the version gap that caused it. `supports_jsonb_subscript` claimed 14.0+ for a capability PostgreSQL does not have at all (SQL subscripting covers arrays and composites; jsonb is read with `->`/`->>` or the SQL/JSON path language) and now answers `False` unconditionally, and `supports_json_arrow_operators` returned `True` unconditionally, advertising on 9.2 — where the operators do not parse — an arrow the core formatter would happily render; it is now gated at 9.3. The four version floors were measured against the scenario servers rather than read off the release notes, because the notes and the servers disagreed about jsonpath: the notes say 13, and 12 has it. `PostgresArrayEnhancedMixin` is now listed before the core `ArrayMixin` and `PostgresUUIDMixin` before the core `UUIDMixin`, where each answers `True` to a question the core answers `False` for; listed in the backend block, the core default was what every caller actually saw and the backend's implementation was unreachable. Each fix is pinned by a test in both directions, so a versioned type cannot again answer yes unconditionally while `text`/`integer`/`varchar` saying yes remains correct on purpose.

changelog.d/62.internal.md

@@ -0,0 +1 @@
+Make Python 3.15 a required CI leg instead of an experimental one: `test-python-315` replaces the `if: ${{ false }}` / `continue-on-error` `test-experimental-python-315` job that ran a 3.15.0-rc.2 pair. The free-threaded 3.15 leg is dropped because psycopg does not support free-threaded builds (psycopg/psycopg#1095), but the free-threaded marker stays in the job *name*, since two legs named alike is how a GIL pass and a free-threaded failure came to look like one job that both passed and failed. The benchmark job moves to Python 3.15 / PostgreSQL 18. `requirements-dev.txt` pins pydantic 2.12.2 / pydantic-core 2.41.4 for Python 3.13 including the free-threaded 3.13t build, restated in this file because it is installed before the core package and an unpinned `pydantic` there resolves to a pydantic-core whose only cp313t distribution is an sdist that cannot build (PyO3 0.29 refuses the free-threaded build of CPython lower than 3.14). CI installs the core from its `release/v1.0.0.dev30` branch now that the column-layer rebuild has merged, and the testsuite from `test/typed-column-contracts`. The expression round-trip matrix now tells a missing formatter apart from a declined feature — core reports both through the same `UnsupportedFeatureError`, which let a class whose formatter this dialect never mixed in pass as "PostgreSQL lacks the feature" — so the method the dispatch names must be a row in the new `UNMODELLED_FORMATTERS` table with the reason (SQL:2003 PIVOT/UNPIVOT and the TableSource root, today), pinned in both directions. New coverage alongside it: column-type suggestions, cast integration, domain-expression and expression round-trips, extension type execution and operations, JSON path modes, UUID values, and the DDL-source accessor and DDL-object contracts.

changelog.d/62.removed.md

@@ -0,0 +1 @@
+**BREAKING**: Remove `PostgresCharacterVaryingType` from the `rhosocial.activerecord.backend.impl.postgres.expression` namespace and its `__all__`, remove `supports_uuid_generation` from `PostgresUuidOssSupport`/`PostgresUuidOssMixin` (renamed `supports_uuid_ossp_extension`, which asks the narrower question of whether the extension is installed; the merged answer now lives only in core's `UUIDSupport` and is folded by `PostgresUUIDMixin`, which answers True for a 13.0+ built-in with no extension at all), and remove `supports_infinity_numeric_infinity_jsonb` from `PostgresDataTypeMixin`, together with the two probes the core protocol's probe-family retirement asked this backend to stop declaring.

requirements-dev.txt

@@ -53,3 +53,17 @@ rich
 
 # Test timeout plugin for long-running database integration tests
 pytest-timeout
+
+# Python 3.13, including the free-threaded 3.13t build. The python_version
+# marker distinguishes release lines, not GIL configuration, so these two lines
+# pin both 3.13 builds. pydantic-core has no cp313t wheel after 2.41.4, and the
+# newer sdists cannot be built there (PyO3 0.29 refuses the free-threaded build
+# of CPython < 3.14); 2.12.2 is the last pydantic paired with 2.41.4.
+#
+# These mirror the same pins in rhosocial-activerecord's own dependencies, and
+# they have to be restated here because this file is installed before the core
+# package: resolving an unpinned `pydantic` here picks a pydantic-core whose only
+# cp313t distribution is an sdist, and the build then fails on the free-threaded
+# jobs before the core's pin is ever in scope.
+pydantic[email]==2.12.2; python_version == '3.13'
+pydantic-core==2.41.4; python_version == '3.13'

src/rhosocial/activerecord/backend/impl/postgres/backend/base.py

@@ -114,6 +114,7 @@ def _register_postgres_adapters(self):
         from ..adapters.monetary import PostgresMoneyAdapter
         from ..adapters.network_address import PostgresMacaddrAdapter, PostgresMacaddr8Adapter
         from ..adapters.object_identifier import PostgresOidAdapter, PostgresXidAdapter, PostgresTidAdapter
+        from ..adapters.uuid import PostgresUUIDAdapter
         from ..adapters.pg_lsn import PostgresLsnAdapter
         from ..adapters.text_search import PostgresTsVectorAdapter, PostgresTsQueryAdapter
         from ..config import RangeAdapterMode
@@ -133,6 +134,11 @@ def _register_postgres_adapters(self):
             PostgresOidAdapter(),
             PostgresXidAdapter(),
             PostgresTidAdapter(),
+            # Same {uuid.UUID: {str}} pair the core UUIDAdapter registers, so
+            # this needs allow_override (see the loop below). Unlike
+            # PostgresJsonPathAdapter it claims no str->str mapping, so it
+            # cannot shadow the default string handling.
+            PostgresUUIDAdapter(),
             # PostgresJsonPathAdapter(), # Not registered: str->str conflict
             # PostgresXMLAdapter is NOT registered by default
             # due to str->str type pair conflict
@@ -141,7 +147,9 @@ def _register_postgres_adapters(self):
         for adapter in pg_adapters:
             for py_type, db_types in adapter.supported_types.items():
                 for db_type in db_types:
-                    self.adapter_registry.register(adapter, py_type, db_type)
+                    self.adapter_registry.register(
+                        adapter, py_type, db_type, allow_override=True
+                    )
 
         # Register Range adapters based on configuration
         range_mode = getattr(self.config, "range_adapter_mode", RangeAdapterMode.NATIVE)

src/rhosocial/activerecord/backend/impl/postgres/dialect.py

@@ -38,6 +38,7 @@
 
     WindowFunctionMixin,
     JSONMixin,
+    UUIDMixin,
 
     ArrayMixin,
     ExplainMixin,
@@ -60,6 +61,15 @@
     ILIKEMixin,
     ConstraintMixin,
     PartitionMixin,
+    # TRIM/LPAD/RPAD/REPEAT are nodes with default formatters, and PostgreSQL
+    # spells all four natively -- ``TRIM([BOTH|LEADING|TRAILING] [chars] FROM
+    # str)``, ``LPAD``/``RPAD(str, len, padstr)`` and ``REPEAT(str, n)`` -- so
+    # the shared defaults are the answer here and no override is needed.
+    LpadMixin,
+    RepeatMixin,
+    RpadMixin,
+    TrimMixin,
+
     # New Mixins
     PredicateMixin,
     ExpressionMixin,
@@ -116,7 +126,6 @@
     CTESupport,
     FilterClauseSupport,
     WindowFunctionSupport,
-    JSONSupport,
     ReturningSupport,
     AdvancedGroupingSupport,
     ArraySupport,
@@ -160,6 +169,7 @@
     PostgresDataTypeMixin,
     PostgresLogicalReplicationMixin,
     PostgresParallelQueryMixin,
+    PostgresColumnTypeMixin,
     # Per-feature mixins
     PostgresCTEMixin,
     PostgresWindowMixin,
@@ -230,6 +240,7 @@
     PostgresFullTextSearchMixin,
     PostgresRangeTypeMixin,
     PostgresJSONBEnhancedMixin,
+    PostgresUUIDMixin,
     PostgresArrayEnhancedMixin,
     PostgresTypeFormatSupportMixin,
     # DDL/DML operation mixins (new)
@@ -418,6 +429,15 @@ class PostgresDialect(
     PostgresCollationMixin,
     PostgresOrderedSetAggMixin,
     PostgresFeaturesMixin,
+    # TRIM/LPAD/RPAD/REPEAT are nodes with default formatters, and PostgreSQL
+    # spells all four natively: ``TRIM([BOTH|LEADING|TRAILING] [chars] FROM
+    # str)``, ``LPAD``/``RPAD(str, len, padstr)`` and ``REPEAT(str, n)``
+    # (functions-string.html; ``TRIM(BOTH 'xy' FROM 'xyxab')`` live-verified
+    # 2026-10-09 as the set form). No override is needed for any of them.
+    LpadMixin,
+    RepeatMixin,
+    RpadMixin,
+    TrimMixin,
     SQLXMLMixin,
     CollationMixin,
     SetOperationMixin,
@@ -428,7 +448,15 @@ class PostgresDialect(
     WindowFunctionMixin,
     PostgresJSONBEnhancedMixin,
     JSONMixin,
-
+    # PG's UUID answer (version 13+ built-in, or uuid-ossp) must win over
+    # the core UUIDMixin table, which has a single fixed spelling per operation.
+    PostgresUUIDMixin,
+    UUIDMixin,
+
+    # Before ArrayMixin: this answers True where the core answers False, and
+    # it was listed in the backend block further down, so the core default
+    # was what every caller actually saw.
+    PostgresArrayEnhancedMixin,
     ArrayMixin,
     ExplainMixin,
     # Must precede GraphMixin/GraphTableMixin so that the explicit-override
@@ -475,6 +503,12 @@ class PostgresDialect(
     PostgresDataTypeMixin,
     PostgresLogicalReplicationMixin,
     PostgresParallelQueryMixin,
+    # Column types. After the DataType mixin above because that is the
+    # sibling decision -- this one says which operations a value carries, the
+    # other says how it is spelled. Nothing else in the list can answer for a
+    # common Python type, so the MRO leaves the placement free; putting the two
+    # side by side is what makes the separation legible to the next reader.
+    PostgresColumnTypeMixin,
     # Extension feature mixins
     PostgresLtreeMixin,
     PostgresIntarrayMixin,
@@ -516,7 +550,6 @@ class PostgresDialect(
     MultirangeMixin,
     PostgresFullTextSearchMixin,
     PostgresRangeTypeMixin,
-    PostgresArrayEnhancedMixin,
     PostgresTypeFormatSupportMixin,
     # DDL/DML operation mixins (new)
     PostgresExtendedStatisticsMixin,
@@ -544,7 +577,14 @@ class PostgresDialect(
     CTESupport,
     FilterClauseSupport,
     WindowFunctionSupport,
-    JSONSupport,
+    # JSONSupport is not listed: PostgresJSONBEnhancedSupport derives from it
+    # below, and listing a base ahead of its own subclass is not a consistent
+    # MRO. This is the shape every other derived protocol uses here --
+    # PostgresTableSupport(TableSupport), PostgresIndexSupport(IndexSupport) --
+    # and JSONSupport still reaches the MRO, so isinstance(dialect, JSONSupport)
+    # is unchanged. Listing both, while each was standalone, also put
+    # supports_json_type in the MRO four times: two abstract protocol
+    # declarations, the core mixin default, and the one implementation.
     ReturningSupport,
     AdvancedGroupingSupport,
     ArraySupport,