perf: post-ingest async embedding overwhelms Ollama under concurrent load

[salishforge]

Problem

When EMBEDDING_PROVIDER=ollama and multiple concurrent add() calls are made, postIngestAnalysis() fires one embed() call per add(). Under benchmark load (150+ concurrent adds), Ollama's embedding endpoint times out with HeadersTimeoutError.

Root Cause

postIngestAnalysis() is fire-and-forget async. With concurrent ingestion, hundreds of embedding requests hit Ollama simultaneously. The local nomic-embed-text model can only process ~2 req/s, causing cascading timeouts.

Impact

• Benchmark with per-session storage (CONSOLIDATION_INNER_BATCH_SIZE=1) + embeddings fails under load
• Individual add() calls still work (the error is non-blocking)
• Consolidation embedding (embedBatch) also competes for Ollama capacity

Proposed Fix

1. Add a concurrency limiter on embedding calls (max 2-3 concurrent Ollama requests)
2. Queue post-ingest analysis instead of fire-and-forget
3. Or: skip postIngestAnalysis when consolidation will embed anyway (avoid double-embedding)