Observability

Both the API and the web app speak OpenTelemetry (docs/observability.md). compose.yaml runs grafana/otel-lgtm: one container with the OTLP collector, Prometheus, Tempo, Pyroscope, Loki and Grafana (http://localhost:3100). Nothing is exported until OTEL_EXPORTER_OTLP_ENDPOINT is set.

• apps/api/src/instrumentation.ts is the first import of main.ts: the Node SDK must patch http, express, pg and the AWS SDK before Nest loads them. Filesystem spans are off, since one pack operation reads thousands of files. The SDK flushes on SIGTERM.
• Tempo's metrics generator turns spans into traces_spanmetrics_*, so latency percentiles exist for every route without code.
• apps/api/src/lib/metrics.ts holds application counters, recorded at the one commit point so each push counts once: ghost_pushes_total, ghost_push_bytes_total, ghost_ref_updates_total. Instance totals (repositories, users, issues…) are observed by AppStatsService.
• apps/delivery exports nothing itself; its numbers come from its tables, so they exist while delivery is down, which is when they matter.
• Pyroscope receives continuous profiles when PYROSCOPE_SERVER_ADDRESS is set.

Logs: Nest's logger, plus morgan request logs in main.ts. The git code logs at debug for each spooled body and each materialization.