perf: Use INSERT RETURNING for all PostgreSQL create operations

[smoreinis]

Summary

Optimize all PostgreSQL create operations by using PostgreSQL's RETURNING clause to get auto-generated fields in a single query instead of two.

Changes

1. Event Repository - Use INSERT RETURNING to get auto-generated sequence_id in single query
2. Base PostgresCRUDRepository - Apply same optimization to all PostgreSQL repositories (agents, spans, tasks, etc.)

Before (2 queries per create)

session.add(orm)
await session.commit()
await session.refresh(orm)  # Extra SELECT to get auto-generated fields

After (1 query per create)

# Return explicit columns to avoid lazy-loading relationships
stmt = insert(self.orm).values(**values).returning(*self.orm.__table__.columns)
result = await session.execute(stmt)
row = result.one()
await session.commit()
return self.entity.model_validate(dict(row._mapping))

Technical Details

• Uses returning(*columns) instead of returning(ORM) to avoid lazy-loading relationships on returned ORM objects
• Excludes None values from INSERT to preserve server defaults (e.g., created_at)
• Works correctly with all ORM models including those with relationships (AgentORM, TaskORM)

Benchmark Results (Events)

Tested with 200 iterations on docker-to-docker network:

Metric	OLD (add+refresh)	NEW (RETURNING)	Improvement
Mean latency	0.812ms	0.684ms	15.8% faster
Median	0.780ms	0.644ms	17.4% faster
P95	1.050ms	0.887ms	15.5% faster
P99	1.780ms	1.009ms	43.3% faster
Throughput	1,231 events/sec	1,462 events/sec	+231 events/sec

Note: Docker-to-docker latency is ~0.1ms. In production with typical network latency (1-5ms), the improvement would be more significant since we're eliminating an entire network round-trip per create operation.

Test plan

• All 15 repository tests pass
• All 184 unit tests pass
• Verified auto-generated fields (sequence_id, created_at) correctly populated
• Tested with ORM models that have relationships (AgentORM with tasks relationship)
• Benchmarked locally with 200 iterations