fix(core): preserve repeated default project selection

[phernandez]

Why

Repeatedly choosing the project that is already the default can leave a workspace with no default project. The API still returns success, but subsequent project discovery and every project-scoped MCP tool fail because no default can be resolved.

A production read-only investigation found repeated successful set-default requests followed by a database state where every project row had is_default = NULL.

What Changed

• Make ProjectRepository.set_as_default idempotent when the selected project is already the default.
• Add a regression test that verifies the committed state from a new database session.

Implementation Details

• Validate and load the selected project before changing any default flags.
• Clear the default flag only from other project rows.
• Preserve the selected row so SQLAlchemy cannot lose the update when its identity-map value was already True before the bulk update.
• Keep the existing repository and API contracts unchanged; no schema migration is required.

Testing

Automated:

• uv run pytest -q tests/repository/test_project_repository.py::test_set_as_default_keeps_existing_default — reproduced the failure before the fix and passes after it.
• uv run pytest -q tests/repository/test_project_repository.py — 14 passed with SQLite.
• BASIC_MEMORY_TEST_POSTGRES=1 uv run pytest -q tests/repository/test_project_repository.py::test_set_as_default_keeps_existing_default — passed with PostgreSQL.
• just fast-check — passed (Ruff formatting/lint and type checking).

Manual:

• Correlated repeated successful set-default requests with the missing-default database state using production read-only telemetry and database inspection.

Risks / Follow-ups

• The change is limited to default-project metadata and has no migration or public API impact.
• Basic Memory Cloud must update its Core revision after this PR merges before hosted behavior changes.
• Existing affected workspaces require a separate recovery action; this PR does not mutate production data.