ntpstats 2.16.0

Research bench v2, part 1 (#29). The new modules are provisional (not yet in ntpstats.api) until
2.17 adds the simulator exchange formats.

Added

• Delay traces (ntpstats.trace, ntpstats trace): per-direction one-way delays extracted
  from any two-way exchanges.
  • Inputs: NTP and PTP captures, chrony measurements.log, ntpd peerstats/rawstats, and
    simulator output.
  • The clock offset between the two ends is removed (--detrend floor|linear|none).
  • The asymmetry is an explicit assumption (--asymmetry), since two-way timestamps cannot
    measure it.
  • ntpstats trace reports floors, PDV percentiles, loss and the correlation of the two
    directions; --csv exports the trace.
• Trace replay in the bench: ntpstats bench trace:FILE, or a [trace] table in a scenario
  file.
  • Modes: replay plays the trace in order; bootstrap draws random blocks, keeping the
    short-term correlation and both directions together.
  • scale multiplies the queueing.
  • Checked by a round trip: delays extracted from a simulated network and replayed give the
    same estimator scores as the model they came from.
• PTP servos and boundary-clock chains (ntpstats.ptpsim, ntpstats chain): a grandmaster
  and N boundary clocks, simulated with ground truth.
  • Each node has its own oscillator, the E2E delay mechanism with a moving-median filter, and
    linuxptp's PI servo (with its default gains) or an adaptive linear-regression servo.
  • Links have asymmetry, timestamp noise, and modelled or replayed PDV.
  • Time-error metrics are given per node and per hop: max|TE|, |cTE|, dTE_L MTIE, dTE_H.
  • Each hop is checked against the commonly quoted G.8273.2 T-BC class limits (A/B/C) or your
    own limits, and the end of the chain against a budget (default 1.1 µs); exit code 3 on
    failure.
  • --kp/--ki tune the servo. The warm-up follows the servo's lock time, and the command
    warns when a run is too short to lock.
• New notebook: a PTP chain against a time-error budget. With linuxptp's default gains, dynamic
  time error grows much faster than the number of hops (gain peaking): about 50 ns at 10 hops,
  about 600 ns at 20. A narrower loop keeps 20 hops near 25 ns. The bench notebook now also
  replays a real log.

Fixed

• Noise identification no longer divides by zero on perfectly alternating or noise-free data (lag-1
  autocorrelation of −1, the white-PM limit).
• Commands piped into head and similar tools stop quietly instead of printing a
  BrokenPipeError traceback.