Keep Float(asdecimal=True) result conversion with the ibm_db driver; declare DOUBLE as Float

[aminghadersohi]

Problem

Float(asdecimal=True) returns float results, not Decimal, with the ibm_db dialect:

from sqlalchemy import Float
from ibm_db_sa.ibm_db import DB2Dialect_ibm_db

d = DB2Dialect_ibm_db()
p = Float(asdecimal=True).dialect_impl(d).result_processor(d, None)
print(type(p(1.25) if p else 1.25))  # <class 'float'>; expected Decimal

DB2Dialect_ibm_db.colspecs maps Numeric to _IBM_Numeric_ibm_db. Float subclasses Numeric and has no entry of its own, so SQLAlchemy's adapt_type adapts every Float (including REAL, DOUBLE_PRECISION and Float(precision=...)) to _IBM_Numeric_ibm_db. That type returns no result processor when asdecimal is true, so the DBAPI's float passes through and the asdecimal=True contract is broken.

Related: ibm_db_sa.base.DOUBLE subclassed Numeric, so reflected DOUBLE columns declared asdecimal=True and python_type Decimal, although DB2 DOUBLE is binary floating point and the DBAPI returns float.

Fix

Add Float: Float to the ibm_db dialect's colspecs. adapt_type then finds Float first in the MRO and keeps the original type, so SQLAlchemy's own Float.result_processor applies. Numeric/DECIMAL still adapt to _IBM_Numeric_ibm_db; nothing else changes. DDL is unaffected (compilation uses the declared type). This is independent of #201 and works with or without it.

Second commit (2e6f403): DOUBLE is now based on Float (asdecimal=False, python_type float). Returned values don't change and DDL still renders DOUBLE. Compatibility note: the Float constructor has no scale argument, so a caller passing DOUBLE(precision, scale) positionally would need adjusting; nothing in this repository does that.

Regression tests

test/test_float.py, same style as the existing tests: connectionless result-processor checks for Float(asdecimal=True), REAL(asdecimal=True) and default Float(), plus live round trips for Float(asdecimal=True) and Float(), comparing values and Python types exactly (no tolerances). For DOUBLE: test_double_is_float (declaration, compilation and result processing) and a live test_reflected_double (reflected type is DOUBLE, asdecimal is False, python_type is float, values/types round-trip as float).

First commit (Float colspec) — same test file, Python, SQLAlchemy, DBAPI and server; only ibm_db_sa/ibm_db.py differs between upstream master 7f3866e3a827924055323fd5f10ac40b7a9fb2a7 and this patch:

Tests	Before	After
Connectionless	2 FAIL / 1 PASS	3 PASS
Live DB2	1 FAIL / 1 PASS	2 PASS
Combined	3 FAIL / 2 PASS	5 PASS / 0 FAIL / 0 SKIP

The failures are type mismatches (float where Decimal is expected); the default Float() tests pass before and after, confirming unchanged behaviour there.

Second commit (DOUBLE) — only ibm_db_sa/base.py differing, full test/test_float.py: 2 FAIL / 5 PASS before, 7 PASS after. Both failures are the declaration assertions; value assertions pass before and after. The live reflection test passes the current schema explicitly, because default_schema_name is '' on current master (fixed separately in #203).

python -m pytest -p sqlalchemy.testing.plugin.pytestplugin --db sqlite -q test/test_float.py::TestFloatResults
python -m pytest -p sqlalchemy.testing.plugin.pytestplugin --dburi "$DB2_TEST_URL" -q test/test_float.py

Validation: Python 3.11, SQLAlchemy 2.0.52, ibm-db 3.2.3, pytest 9.1.1, local throwaway DB2 Community Edition 11.5.9.0 container. The connectionless tests also pass on SQLAlchemy 1.4.54 (the live class is skipped there on sqlite). Ruff 0.5.0 check/format on the new test file and git diff --check pass. The entire legacy upstream suite was not run.

[aminghadersohi]

Added a second commit, 2e6f403 ("Declare DOUBLE as a Float type"), for a related declaration mismatch.

ibm_db_sa.base.DOUBLE subclassed Numeric, so reflected DOUBLE columns declared asdecimal=True and python_type Decimal, although DB2 DOUBLE is binary floating point and the DBAPI returns float. DOUBLE is now based on Float (asdecimal=False, python_type float). Returned values don't change and DDL still renders DOUBLE. One compatibility note: the Float constructor has no scale argument, so any caller passing DOUBLE(precision, scale) positionally would need adjusting. Nothing in this repository does that.

Tests: test_double_is_float (declaration, compilation and result processing) and a live test_reflected_double (reflected type is DOUBLE, asdecimal is False, python_type is float, and values/types round-trip as float). Same local DB2 11.5.9.0 server, only ibm_db_sa/base.py differing: 2 FAIL / 5 PASS before, 7 PASS after. Both failures are the declaration assertions; value assertions pass before and after. Connectionless tests also pass on SQLAlchemy 1.4.54.

Separately observed on current master and not changed here: inspect(connection).default_schema_name is '', so get_columns(table) without an explicit schema returns no columns. The live test passes the current schema explicitly.