Fix migration completion checks

[mameikagou]

What changed

• Give each schema migration its own completion column (user_id for 1.2.0 and project_id for 1.3.0).
• Check the real schema even when schema_version claims the migration already ran, so databases incorrectly marked as 1.3.0 can self-repair.
• Apply the same completion logic to SQLite and PostgreSQL migrations.
• Add SQLite regression coverage for the corrupted 1.3.0 state and repeated, idempotent execution.

Why

The migration runner used memories.user_id as the applied check for every migration. Once 1.2.0 added that column, 1.3.0 was skipped and the missing project_id columns were still recorded as applied.

The version record alone cannot repair databases that already hit this bug, so the runner now also checks the migration-specific completion column.

Impact

Existing databases marked as 1.3.0 but missing project isolation columns are repaired on the next migration run. Already-correct databases remain unchanged.

Validation

• npm run typecheck
• npx vitest run tests/migrate.test.ts --maxWorkers=1 --no-file-parallelism
• npm test (42 tests passed)

Closes #196

[nullure]

LGTM