Fix: Resolve OOM DoS Vulnerability and Optimize FAISS I/O Caching (Clean)

[Adar5]

Per Leo's suggestion, I am opening this clean branch to isolate only the necessary fixes for the OOM vulnerability and FAISS caching, leaving behind any unrelated noisy commits from the previous PR.

Key Changes:

OOM DoS Mitigation (Defense-in-Depth):

Implemented a strict 2000-character truncation shield on all input gateways (Standard Chat, WebSockets, and File Upload endpoints) to prevent massive string payloads from exhausting RAM.

Migrated from ConversationBufferMemory to ConversationBufferWindowMemory(k=10) to bound the session history, preventing long conversations from causing gradual memory leaks while preserving the secure persistence feature.

FAISS I/O Caching:

Applied @lru_cache(maxsize=1) strictly to the FAISS index and metadata loaders to prevent redundant disk reads.

Refactored dynamic loggers out of the RAG cache boundary to resolve unhashable type errors during inference.

Test Suite Modernization:

Consolidated pytest_plugins from nested conftest.py files into a single root-level conftest.py to fix collection errors in newer Pytest versions (8.4.1+).

Testing:

Successfully survived local stress testing using extreme multi-megabyte payload scripts without causing Uvicorn crashes or memory exhaustion.

Full test suite passing locally (161/161 passed).