Major Refactor: Killing Memory Chunker

The memory chunker worker has been removed, significantly reducing codebase size by approximately 3,000 lines.

Traits extraction is now handled via a dedicated asynchronous LLM call using a simplified prompt, utilizing a smaller 4B model.

Communication style metrics are calculated deterministically by the StyleMetricsService, requiring zero LLM usage and taking about 1ms.

Gists and Facts have been eliminated from the pipeline; raw conversation turns in working memory now serve as the replacement.

Emotion and Scope extraction are simplified; emotion is extracted independently by the episode worker, and boundary detection plus idle signals replace scope-based consolidation triggers.

The trait service has been streamlined, dropping parameters like source and is_literal, and collapsing trait categories from eight to three.

The episodic memory pipeline now triggers observation based on a minimum turn count of five, instead of relying on signal density from memory chunks.

This refactor updated 71 files across various components and maintained 1549 passing tests.

• Migration 002 fixes schema matching issues on fresh installs, resolving sqlite3.OperationalError.

• Migration 007 drops unused semantic_schemas and cognitive_reflexes.

• Removed dead table definitions and unused columns from schema.sql.

• Added bounds checking on label/logit array indexing in onnx_inference_service.py.

• Added cognitive_drift:* glob pattern to ensure legacy MemoryStore keys are cleaned during data deletion.

• Added embeddings mock to health check test for /system/ready endpoint.