AI Insights: reliable Ollama-backed facet extraction for deck analytics

[bdougie]

Summary

The AI Insights panel in deck --web analytics is currently disabled. The Ollama-backed facet extraction has reliability and performance issues that need to be resolved before re-enabling.

Problems

• Ollama responses often fail JSON parsing (invalid character 'O' looking for beginning of value) even with format: "json" set
• Backfilling all sessions blocks the /api/facets endpoint, making the analytics page load extremely slow
• Auto-detecting Ollama on startup adds latency even when insights aren't needed

Current State

• UI is commented out in deck.js (analyticsInsightsEl.hidden = true)
• Server-side facet endpoints still exist and work with --insights flag
• facets_ollama.go has format: "json" set but models still return non-JSON

Proposed Work

• Investigate async/background backfill so it doesn't block page load
• Add retry logic or better prompt engineering for JSON extraction
• Consider streaming partial results to the UI as sessions are processed
• Re-enable the insights UI once extraction is reliable

[linear]

PCC-108