feat(retain): add affect recognition metadata

[Musicaldg]

Summary

• add opt-in sentiment/emotion recognition to the Retain extraction schema
• propagate versioned affect metadata through the HMS ingestion pipeline and persist it on memory units
• add PostgreSQL and Oracle migrations/write paths while leaving Recall unchanged
• document deployment configuration with safe defaults and fail-soft parsing

Design

• taxonomy: positive/negative/neutral sentiment plus joy, sadness, anger, fear, surprise, disgust, and neutral emotion
• intensity is constrained to 0.0–1.0
• classification uses the existing structured Retain LLM call, avoiding an extra provider round trip
• feature defaults off and invalid model output degrades to NULL rather than failing Retain

Validation

• Retain offline CI-equivalent suite: 316 passed, 3 skipped
• focused affect, Alembic DAG, and database abstraction suite: 94 passed
• related extraction, projection, labels, and batch regression suite: 86 passed (DB-dependent cases separately covered by deployed smoke test)
• deployed PostgreSQL migration advanced to s7t8u9v0w1x2
• real fact_storage -> PostgreSQLOps write/read smoke test preserved the affect JSON and cleaned up the test row
• deployed API health check: healthy, database connected
• Ruff, format, compileall, Docker Compose config, and diff checks passed