Add observability signal catalog (operational + business metrics)

[rororowyourboat]

Context

From the system design audit, observability is stated as a goal (OTLP trace proxy exists) but there is no unified signal catalog defining what is measured and what is excluded.

Design Decision (from spec)

Operational + business metrics. No conversation content in traces.

Operational Metrics

• event.persisted.count, event.persisted.latency_ms
• projection.rebuild.duration_ms
• rpc.request.count (by method), rpc.request.latency_ms, rpc.error.count
• ws.connection.count, ws.reconnect.count
• canvas.snapshot.save.latency_ms
• pty.session.count

Business Metrics

• turn.completed.count (by provider, model), turn.duration_ms
• turn.interrupted.count, turn.error.count
• thread.created.count
• approval.requested.count, approval.decision.count (by decision type)
• git.action.count (by type), git.action.duration_ms
• checkpoint.revert.count, diff.viewed.count

Content Exclusions (Mandatory)

Never include: message.text, attachment content, diff blobs, terminal I/O, file contents, provider API keys.

Proposed Changes

1. Define signal names, types, dimensions, and collection points in a canonical doc
2. Instrument key paths: event persistence, RPC dispatch, turn lifecycle, git actions
3. Ensure OTLP trace export respects content exclusions
4. Add counters/timers at the collection points listed above

Acceptance Criteria

• Signal catalog documented
• At least operational metrics instrumented at key collection points
• Content exclusion rule enforced (no message text/files/diffs in traces)

References

• System Design Spec: .plans/21-system-design-spec.md Section 8 (Quality Attributes)
• Audit finding: "Observability is stated as a goal but has no signal catalog"

[rororowyourboat]

Implemented in commit 5d44e40 — Signal catalog at .plans/22-observability-signal-catalog.md, instrumentation markers added