fix(postgresql): seed Npgsql's global type mapper before the test suite runs (weasel#398)

[jeremydmiller]

Closes #398.

TL;DR

The parameter type is not silently wrong. add_first_parameter asserts on Npgsql's inference, and that inference depends on process-global state that another test happens to initialize first. In a cold process it fails 100% of the time.

What actually happens

command.AddParameter("a") passes dbType: null, so DatabaseProvider.AddParameter never calls SetParameterType:

parameter.Value = value ?? DBNull.Value;
if (dbType.HasValue)   // <-- false
{
    SetParameterType(parameter, dbType.Value);
}

PostgresqlProvider.determineParameterType, ResolveNpgsqlDbType and ParameterTypeMemo are not on this path at all. So the fix suggested in the issue — seeding ParameterTypeMemo in storeMappings() — would not have changed this test's outcome. Neither would anything about NpgsqlTypeMapper.Mappings; that static Cache is only read by Weasel's own resolution, which never runs here.

_npgsqlDbType is left null, so the param.NpgsqlDbType getter infers from Value via Npgsql's process-global type mapper. That mapper stays empty — reporting Unknown for every value — until the process constructs its first NpgsqlDataSource or NpgsqlConnection with a connection string. Measured, one fresh process per row:

what ran first in the process	new NpgsqlCommand().AddParameter("a").NpgsqlDbType
nothing	Unknown
new NpgsqlConnection() (parameterless)	Unknown
new NpgsqlConnectionStringBuilder(cs)	Unknown
new NpgsqlConnection(cs)	Text
new NpgsqlDataSourceBuilder(cs) / .Build()	Text
new NpgsqlDataSourceBuilder()	Text

So the test passes only when some other test in the same process built a data source first. Under xUnit v3's parallel collections that is a race — hence 3 of 4 jobs failing and a clean re-run going green. It did not reproduce locally because a local full-suite run almost always wins that race.

This is a test-visibility artifact, not a wire-level defect: with _npgsqlDbType unset, Npgsql resolves the type at execution time against the real connection, and a command cannot execute without a connection existing.

The fix

Seed the mapper once from the existing module initializer, before any test runs. new NpgsqlDataSourceBuilder() is enough — no sockets, no connection string.

Why not make AddParameter set the type

That was the tempting fix (AddNamedParameter already does it) and it is a regression. Weasel's mapping is per-type; Npgsql's inference is per-value. Across the common CLR types DateTime is the only divergence, and it is a breaking one:

value	Npgsql infers	Weasel would set
DateTime.UtcNow	TimestampTz	Timestamp
DateTime.Now	Timestamp	Timestamp

Verified against a live Postgres:

untyped (today's AddParameter behavior)     -> OK
NpgsqlDbType=Timestamp (Weasel's mapping)   -> ArgumentException: Cannot write DateTime
    with Kind=UTC to PostgreSQL type 'timestamp without time zone' ...

All other checked types agree exactly: string, char, int, long, short, double, float, decimal, bool, Guid, DateTimeOffset, TimeSpan, DateOnly, TimeOnly, byte[], string[], int[], List<int>, IPAddress.

Two new tests pin this so the trap is not walked into later:

• add_parameter_honors_an_explicit_type — the path Weasel controls, independent of global mapper state.
• add_parameter_without_a_type_defers_to_npgsql_rather_than_weasels_mapping — fails if AddParameter ever starts stamping Weasel's mapping onto untyped parameters.

Verification

Counterfactual, with the warm-up commented out, cold process:

Failed  CommandExtensionsTests.add_first_parameter
Failed  CommandExtensionsTests.add_parameter_without_a_type_defers_to_npgsql_rather_than_weasels_mapping

— the exact CI failure. With the warm-out restored, green, including 5 consecutive cold-process runs of that class alone.

Full suite, all four CI matrix legs (net9.0/net10.0 x USE_CASE_SENSITIVE_QUALIFIED_NAMES true/false): 786 passed, 0 failed, 3 skipped each.

No product code changed.

Follow-ups, deliberately not in this PR

1. AddParameter vs AddNamedParameter disagree on untyped values — the former defers to the driver, the latter calls ToParameterType. Worth reconciling, but it needs the DateTime-Kind question answered first.
2. AddNamedParameter(cmd, name, DateTime.UtcNow) already sets Timestamp and so already throws on write. Pre-existing, separate from this flake.
3. NpgsqlTypeMapper.Mappings really is a mutable static Cache enumerated without synchronization, and PostgresqlProviderTests writes to it while other collections read. Not the cause here, but still a genuine hazard for anything that does go through Weasel's resolution.

🤖 Generated with Claude Code