feat(core): Implement Sandboxed Execution (ENH-001)

[nadeem4]

This pull request introduces a robust sandboxing architecture for SQL execution and schema indexing, significantly improving crash isolation and reliability for the NL2SQL engine. The main changes include the addition of a global process pool manager for sandboxed execution, updates to the executor node to offload query execution to isolated processes, and documentation updates to reflect the new architecture and its operational invariants.

Sandboxed Execution Architecture

• Added SandboxManager in nl2sql/common/sandbox.py to manage two global process pools: one for latency-sensitive SQL execution and another for high-throughput schema indexing. This ensures that unsafe operations (driver segfaults, OOMs) are isolated from the main application.
• Updated settings in nl2sql/common/settings.py to configure the number of workers for both execution and indexing pools via environment variables.

Executor Node Refactor

• Refactored the ExecutorNode in nl2sql/pipeline/nodes/executor/node.py to offload SQL execution to the sandboxed process pool. This includes re-instantiating adapters inside worker processes and handling worker crashes (segfaults, OOMs) with improved error reporting and safeguards. [1] [2] [3] [4]

Documentation and Architectural Updates

• Revised README.md to describe the new multi-plane architecture (Control, Security, Data), sandboxed execution, and system invariants such as "No Unvalidated SQL" and "Zero Shared State".
• Updated technical documentation in docs/core/architecture.md, docs/core/nodes.md, and docs/core/indexing.md to detail the sandboxed execution model, process pool isolation, and how indexing and query execution are offloaded. [1] [2] [3]
• Clarified the phased rollout of sandboxing in the ADR for sandboxed execution (ADR-001_sandboxed_execution.md).

Adapter SDK Update

• Minor update to the SQLAlchemy adapter to store connection arguments explicitly, supporting better process isolation.