feat: add pluggable chat storage with optional SQLite persistence and session-scoped caching

[tisha-varma]

Summary

Closes #60

Adds a modular storage abstraction for chat history with optional SQLite persistence and an optional session-scoped response cache to reduce redundant LLM calls.

Both features are opt-in. Default behavior remains fully in-memory and unchanged.

No new dependencies — implementation uses only Python stdlib (sqlite3, hashlib, collections.OrderedDict, threading).

---

What Changed

New Files

File	Purpose
backend/chat_storage.py	ChatStorage abstraction + InMemoryChatStorage (default) + SQLiteChatStorage (opt-in, WAL-enabled)
backend/response_cache.py	ResponseCache with TTL expiry and LRU eviction (session-scoped)
tests/test_chat_storage.py	17 unit tests for both storage backends + factory
tests/test_response_cache.py	12 unit tests for cache (hit/miss, TTL, LRU eviction, session scoping)

---

Modified Files

File	Change
backend/agents.py	NeuroscienceAssistant now uses pluggable storage and cache instead of raw dicts
backend/main.py	Added lightweight observability endpoints: GET /api/sessions, GET /api/cache/stats, POST /api/cache/clear
.env.template	Added ENABLE_PERSISTENCE, ENABLE_CACHE, and tuning variables
.gitignore	Excludes generated chat_history.db

---

Design Notes

• Backward-compatible — NeuroscienceAssistant() works unchanged. handle_chat signature remains -> str. Default behavior is still in-memory.
• Graceful fallback — If SQLite fails (e.g., read-only filesystem), the factory falls back to in-memory storage with a warning.
• Session-scoped cache — Cache keys include session_id to prevent cross-session response leakage.
• Large data remains in-memory — all_results lists are kept in _results_buffer to avoid database bloat.
• Thread-safe SQLite — Uses a persistent connection (check_same_thread=False) guarded by a lock.

---

Configuration

# Optional features (both default to false)
ENABLE_PERSISTENCE=true
ENABLE_CACHE=true

# Storage configuration
CHAT_DB_PATH=./chat_history.db

# Cache tuning
CACHE_MAX_SIZE=128
CACHE_TTL_SECONDS=3600

---

Testing

• 29 unit tests, all passing
• Covers storage backends, cross-instance persistence, fallback behavior
• Covers cache TTL, LRU eviction, session scoping
• Suite runtime: ~0.27s
• No flaky tests

---

All features are opt-in and can be toggled independently.

Happy to adjust scope or split changes based on feedback.