v0.63.1

Fixed — telemetry no longer probes the GCE metadata server at startup (ADR-0068)

• With [telemetry].backend = "gcp", startup intermittently paid
  4.5–7.2 s of silence before the banner or the one-shot stdin read
  (measured 6 of 24 runs; the rest paid 0 ms). Traced to Cloud
  Logging's client.Logger, which classifies the host by fetching
  from the GCE metadata server: on a Mac the link-local fetch blocks
  on the kernel's ARP probe, its 2 s dial timeout is retried as
  transient, and the cost depends on whether the kernel's negative
  neighbour entry is still fresh. ADC and genai.NewClient were never
  involved
• The exporter now declares the global monitored resource labelled
  with project_id — exactly what the detection fell back to on this
  platform — so records are unchanged and construction touches no
  network. Measured after the fix: 12 of 12 runs reach the stdin read
  within 16 ms
• No wait notice is added: the wait was a library auto-detection, not
  a contract (ADR-0033 §2 / ADR-0067 announce deliberate waits). A
  test pins the fix by counting hits on a fake GCE_METADATA_HOST