Periodic guard for the NIF layer. Add it to your supervision tree:

children = [
  {HexagonTpu.Monitor, interval_ms: 30_000}
]

Every interval it:

• emits [:hexagon_tpu, :stats] telemetry with the native counters, memory snapshot and dirty-scheduler utilization (attach your metrics reporter to ship these)
• logs a warning when live QNN resource counts keep growing across growth_checks consecutive samples (leak suspicion) or when dirty scheduler utilization exceeds dirty_high_water (default 0.9)

Options:

• :interval_ms — sample period (default 30_000)
• :dirty_sample_ms — microstate-accounting window (default 500)
• :growth_checks — consecutive growing samples before warning (default 3)
• :dirty_high_water — utilization warning threshold (default 0.9)