Add client-perf-diagnosis skill

[habdelra]

Background

The client performance telemetry adds an always-on browser instrument that beacons six event types to Loki and a Grafana dashboard, so real client behavior is measurable and attributable to a specific user or session. This adds the operator-facing counterpart: a diagnostic skill that turns a field complaint into a measured investigation against that telemetry.

The motivating shape of the complaint is concrete — "user XYZ said the app was slow loading their card" — and the whole reason attribution is in the pipeline is to answer it. So the skill leads with the user-attributed drill-down: pivot to the account's matrix_user_id, pin the session, find the slow card-load (the "Loading card…" window) and its settle_ms, then spider into the sibling events in that same session that explain the cost.

What the skill covers

A mode per investigation, all sharing the matrix_user_id / session_id drill-down:

• A user reports a slow card load — the hub mode; from the pinned session, spider into slow server round-trips, a main-thread freeze, heavy deserialization, rebuild churn, or index-event churn.
• A browser froze — read the wedge events and their fn @ url:char blocking-frame breadcrumb (source-map-resolvable for minified builds), plus profiler stacks where sampled.
• Slow server round-trips — server-request latency by endpoint, joined to the realm-server's own realm:requests / realm:search-timing stage breakdown via the logging correlation id.
• Heavy deserialization, rebuild churn (grouped by the module that forced it), and realm index-event write-burst churn — framed as the causal upstream that explains why a rebuild fired or a linked card reloaded.

It documents the pipeline, the Loki label-vs-JSON-field split, the six event schemas, the working LogQL for each mode (lifted from the dashboard so they match the instrument), the local-dev bring-up, and the seam to indexing-diagnostics for the server-side half of a request.

Notes

• Stacked on the telemetry instrument + dashboard change; this branch adds only the skill file. Retarget to main once the base merges.
• The verified half — the mechanism, event glossary, dashboard, and query building blocks — is grounded in the local pipeline. The symptom interpretation is written as "how to investigate" with a calibrate-against-the-realm's-baseline caveat rather than fixed thresholds, so it stays honest as real field data accrues.

🤖 Generated with Claude Code