fix: resolve broken imports, hardcoded credentials, GPU-only crashes, and silent errors across API and backend services

[Copilot]

Multiple services in SentinelAI failed at startup or import time due to missing modules, hardcoded credentials, and unconditional CUDA assumptions. All existing tests were also broken due to a wrong import path in conftest.py.

Broken imports / missing modules

• api/main.py and api/inference.py both imported from a non-existent core.* namespace — created api/core/ package with SentinelModel (2-layer MLP) and fixed all import paths to use api.core.*
• Added pytest.ini with pythonpath = . so api.* resolves during pytest runs from repo root
• Fixed tests/conftest.py: from app.main import app → from api.main import app (all 4 tests were failing on collection)

Auth hardening

• api/auth.py had literal admin/admin check — replaced with API_USERNAME / API_PASSWORD env vars; endpoint returns 500 when API_PASSWORD is unset rather than silently accepting any login

GPU / runtime crashes

• backend/app/main.py: .cuda() called unconditionally → replaced with torch.device("cuda" if ... else "cpu") pattern
• backend/app/main.py: mlflow.start_run() executed at module level → wrapped in try/except so import succeeds when MLflow is unreachable
• api/routes/inference.py: LLM loaded synchronously at import time → lazy-loaded on first request, model name now read from LLM_MODEL_NAME env var, returns 503 on load failure, CPU fallback added

Silent failure → observable errors

• llm-guard/app.py: except Exception: pass on DB write replaced with logger.exception() + conn.rollback()

API completeness

• api/main.py: wired up /infer POST endpoint (with Bearer-token auth dependency + SlowAPI rate limiting) and GET / root route — both were missing, leaving tests hitting 404

Dependencies / lint

• requirements.txt: added httpx (required by FastAPI TestClient) and pydantic
• Fixed 4 pre-existing ruff errors in llm-guard/app.py and streamlit-dashboard/app.py (unused imports, spurious f-string prefix, unused exc binding) that were failing the CI lint job

Original prompt

Create a pull request in Trojan3877/SentinelAI that updates the code files as needed to improve the project.

Requirements:

• Inspect the repository structure and identify outdated, broken, or low-quality code.
• Make targeted updates to code files to improve reliability, maintainability, and correctness.
• Ensure the project runs (or has clear run instructions) after changes.
• Add/adjust any necessary dependency or configuration updates (e.g., requirements.txt / package.json) if needed.
• Keep changes focused and avoid unrelated refactors.
• Add or update a brief section in the README describing what changed and how to run/verify.

Deliverables:

• A PR with clear title and description summarizing changes.
• Updated code files committed to a new branch.
• Any basic tests or smoke checks added or documented if appropriate.

The following is the prior conversation context from the user's chat exploration (may be truncated):

User: https://github.com/Trojan3877/Facial-Emotion-Recognition-System open a PR and update code files

Assistant: I’m ready to proceed with opening the pull request and updating the code in Trojan3877/Facial-Emotion-Recognition-System—please respond to the pending prompt (accept or reject) so I can continue.

Assistant: I've initiated the process to open a pull request and update the code files in the repository Trojan3877/Facial-Emotion-Recognition-System. The Copilot coding agent is now working on your request.

You'll be notified once the task is complete.

This pull request was created from Copilot chat.