[CHORE][PERFORMANCE]: Reduce mcpgateway.main import-time bootstrap cost for test fixtures

[jonpspri]

Problem

Importing mcpgateway.main takes ~2.7–3.0s due to module-level side effects that run at import time. This cost is paid by every test that uses the function-scoped app fixture in tests/conftest.py, making those tests significantly slower than they need to be.

Affected test examples (setup time, not test execution):

• tests/unit/mcpgateway/test_main_extended.py::TestAdminAuthMiddleware::test_admin_auth_trailing_slash_root_path — 2.78s setup
• tests/unit/mcpgateway/routers/test_metrics_maintenance.py::test_metrics_rollup_disabled — 2.73s setup
• tests/unit/mcpgateway/routers/test_well_known_rfc9728.py::TestRFC9728CompliantEndpoint::test_rfc9728_endpoint_oauth_not_enabled — 2.71s setup
• tests/unit/mcpgateway/test_main_extended.py::TestConditionalPaths::test_redis_initialization_path — 2.71s setup

Root cause

mcpgateway/main.py performs substantial work at module import time (lines ~160–227):

1. Database readiness check (wait_for_db_ready) — synchronous retry loop at import time (line 172)
2. Database bootstrap (bootstrap_db()) — runs Alembic migrations, creates tables, creates admin users, sets up RBAC roles (line 178). This is the most expensive step.
3. Redis readiness check (wait_for_redis_ready) — another synchronous retry loop (line 217)
4. Service singleton instantiation — ~10 service singletons created at module level (lines 196–213)
5. Session registry initialization (lines 220–227)
6. Heavy import chain — main.py imports ~50 modules at the top level (lines 30–159), each potentially triggering further import-time work

The function-scoped app fixture in tests/conftest.py does from mcpgateway.main import app on every test invocation. While Python caches module objects, the fixture still creates a fresh temp SQLite DB, patches SessionLocal across 6+ modules, and runs Base.metadata.create_all() per test.

Why we can't simply use module-scoped fixtures

A module-scoped app_with_temp_db fixture already exists, but most tests mutate shared state (DB rows, app.dependency_overrides, patched settings) that would leak between tests. Switching to module scope would require adding explicit cleanup/rollback logic to every test, which is fragile and increases maintenance burden.

Suggested investigation areas

1. Defer bootstrap_db() to lifespan — Move the Alembic migration + admin user creation out of module-level code and into the lifespan() async context manager, which already handles all other service initialization. Test fixtures could then skip it entirely.

2. Lazy service singletons — Replace module-level singleton instantiation (lines 196–213) with lazy initialization patterns (e.g., functools.lru_cache or get_*_service() factory functions that several services already use).

3. Guard DB/Redis readiness checks — wait_for_db_ready and wait_for_redis_ready run unconditionally at import time. These could be deferred to lifespan() or guarded with a SKIP_STARTUP_CHECKS env var for test environments.

4. Reduce SessionLocal patch surface — Currently, SessionLocal must be patched in 6+ modules because each does from mcpgateway.db import SessionLocal at import time, capturing the original reference. Migrating these to import mcpgateway.db as _db_mod + _db_mod.SessionLocal() (as was done for base_service.py in PR [EPIC][A2A]: A2A Protocol v0.3.0 Full Compliance Implementation #3150) would allow a single patch on mcpgateway.db.SessionLocal to propagate everywhere.

5. Lightweight test app factory — Consider a create_test_app() factory that builds a minimal FastAPI app with only the routers/middleware needed for the test, rather than importing the full production app.

Acceptance criteria

• Identify the top contributors to the ~3s import cost (profile with python -X importtime)
• Reduce function-scoped app fixture setup time to <1s without sacrificing test isolation
• No regressions in test pass rate or production startup behavior

[crivetimihai]

Thanks @jonpspri. Agreed that the module-level bootstrap in main.py is costly for test fixtures. We'll look at lazy initialization or deferred bootstrap. Tagging as COULD for Release 1.1.0.