Viewer: add retrieval observability panels on top of #26

[acgxv]

Summary

Build an observability-first extension of the read-only viewer so retrieval quality and cost can be debugged without API spelunking.

Why

Issue #26 defines a read-only viewer. We should make it useful for tuning, not only inspection.

Scope

• Extend Add a read-only web viewer for sessions and traces #26 with search pipeline observability panels:
  • stage latency (expand/embed/retrieve/fuse/rerank/filter)
  • candidate counts per stage
  • rerank delta (before vs after rank movement)
  • top explain factors per selected note
• Add side-by-side drill-down flow:
  • search index -> timeline -> note details -> trace explain
• Add mobile-first layout checks for the inspection workflow.

Acceptance Criteria

• A developer can identify where recall or latency regresses from a single viewer session.
• Viewer displays search traces and explain details without extra API tooling.
• Workflow is usable on desktop and mobile.

Related

• Add a read-only web viewer for sessions and traces #26

[acgxv]

Update: treat this as Phase-2 on top of the Phase-1 inspector in #55

Suggested dependency order

1. Ship the MVP inspector workflow (trace list entry point + drilldown + trajectory + export) as described in Operator observability: memory inspector and diagnostics workflow #55.
2. Then add the Viewer: add retrieval observability panels on top of #26 #41 panels on top once operators can reliably reach the right trace/item.

Panel ideas (starting list)

• Per-stage candidate counts + retention (prefilter -> retrieval -> fusion -> rerank -> final)
• “Why excluded?” buckets for prefilter/filters (where available)
• Rerank delta view (rank movement + score contribution summary)
• Recursive/structured/graph context contribution visibility (bounded, explain-only)