Stabilize Sandman Inference Pipeline & Fix SQLite Race Conditions

Description

The autonomous REM sleep pipeline (Sandman) stalls and occasionally crashes due to VRAM starvation and a race condition during SQLite matrix initialization.

Issues Identified

1. Dynamic Context Scaling: Statically assigning num_ctx: 200000 causes OS VRAM starvation on smaller/medium payloads and massive context ingestion latency.
2. Capability Gap Prompt Thinning: Passing 7500+ paths from the workspace tree into the context per node evaluation causes sequence bottlenecks.
3. HTTP Timeouts: Using native fetch() encounters undocumented 5-minute timeout caps from undici during heavy local LLM processing.
4. SQLite Race Condition: Concurrent async operations during local matrix initialization triggered UNIQUE constraint failed: vector_collections_meta.name.

Proposed Fixes

• Dynamically allocate KV cache with Math.ceil(chars / 3) + 4096.
• Implement strict semantic fuzzy matching to limit directory contexts.
• Swap fetch() for native http/https to enforce explicit 1-hour timeouts for Ollama backend calls.
• Implement INSERT OR IGNORE INTO in SQLiteVectorManager.mjs to handle concurrent discoveries gracefully.