Optimize DuckDB sync pipeline after merge projection work

[marcusschiesser]

Context

Follow-up work from docs/sync-optimization-ideas.md.

Current 10k trace sync timings:

Run	Initial Sync	No-Op Sync
Final optimized merge projection	5.80-5.88s	~22-23ms

Current representative profiled cost shape:

Phase	Time
Object load + OTEL conversion	2.86-2.90s
Raw row normalization	0.28-0.30s
Projection total	~1.70s
Bookkeeping	~0.002s

Tasks

• Pipeline object loading and projection with bounded concurrency while keeping DuckDB writes serialized.
• Parse OTEL resource spans directly into RawEventRow records for sync, skipping the intermediate IngestionEvent[] allocation where possible.
• Build pre-normalized projection rows or deltas during OTEL conversion so projection does not re-normalize event bodies.
• Benchmark vectorized DuckDB bulk loading for raw event storage and observation appends against the current row appender path.
• Remove in-memory eventJson from RawEventRow if no hot compatibility path depends on it.
• Add an explicit empty projection fast path that skips existing-row probes, deletes, and historical replay checks on fresh caches.

Acceptance Criteria

• Each task has before/after timing for the 10k trace perf run.
• pnpm --filter @agentpond/duckdb test passes for each completed task.
• pnpm test passes before closing the ticket.
• Update docs/sync-optimization-ideas.md if implementation results change the recommended order or measured bottlenecks.