fix(#3071): align Sqlite DLQ expiration column with the shared 'expires' name

[jeremydmiller]

Closes #3071.

Diagnosis

The Sqlite DeadLettersTable named the conditional expiration column keep_until (DatabaseConstants.KeepUntil) while every other RDBMS backend — Postgres, SqlServer, MySql, Oracle — names it expires (DatabaseConstants.Expires):

// Wolverine.Postgresql / SqlServer / MySql / Oracle Schema/DeadLettersTable.cs
AddColumn<DateTimeOffset>(DatabaseConstants.Expires).AllowNulls();

// Wolverine.Sqlite Schema/DeadLettersTable.cs (pre-fix)
if (durability.DeadLetterQueueExpirationEnabled)
{
    AddColumn(DatabaseConstants.KeepUntil, "TEXT");   // ← wrong constant
}

The shared DLQ insert path (Wolverine.RDBMS.DatabasePersistence.WriteDeadLetter) and the cleanup query (Wolverine.RDBMS.Durability.DeleteExpiredDeadLetterMessagesOperation) both target DatabaseConstants.Expires:

deadLetterFields += ", " + DatabaseConstants.Expires;
// "delete from {schema}.{DeadLetterTable} where {Expires} < @utcnow"

So a Sqlite host with DeadLetterQueueExpirationEnabled = true provisions a column the shared SQL never writes to and then hits SqliteException: no such column: expires on the periodic cleanup job (and on the next DLQ insert). The reporter calls out that this happens on a fresh database too — confirmed by the regression test.

The fix

Rename the column to expires so it matches the constant the shared SQL uses. One line in Wolverine.Sqlite/Schema/DeadLettersTable.cs. Aligns Sqlite with every other backend.

Regression coverage

SqliteTests/Bug_3071_sqlite_dlq_expiration_creates_expires_column.cs — mirrors the bug report verbatim:

1. dlq_table_has_expires_column_when_expiration_is_enabled — boots a host with PersistMessagesWithSqlite + DeadLetterQueueExpirationEnabled = true + AddResourceSetupOnStartup, then asserts the wolverine_dead_letters table carries expires and not the legacy keep_until. Fails on pre-fix code with the column-name mismatch.
2. dlq_table_has_no_expires_column_when_expiration_is_disabled — regression guard the other way: with expiration off, neither column appears. Locks in the gating so a careless future refactor can't reintroduce the bug from the other side.

Reflects column names via PRAGMA table_info directly rather than going through Weasel's FetchExistingAsync so the assertion shape is independent of any future Weasel column-name normalization choice.

All 142 SqliteTests green (1 pre-existing skip).

Why no migration

The bug never produced a working install — the cleanup job throws on every pass — so there's no existing fleet of Sqlite databases out there with a populated keep_until column to migrate. Anyone who flipped the flag on either ran into the exception immediately or never enabled the feature. Renaming the column is the right end-state and any in-flight database that did happen to provision the wrong-named empty column will simply re-provision the correct one on next resource-setup pass.

🤖 Generated with Claude Code