[cboulay]

Stacked on #18. Everything here was measured against a real Net Amps 500 / GES 500 (firmware 0.4.9). All app changes are gated on the NA500; the NA400 code path is unchanged.

Conformance recorder and NA500 mock

• scripts/ampserver_conformance.py records how an AmpServer behaves on the wire, and diffs two recordings. The recording of the real NA500 is mock/conformance/na500_fw0.4.9.json.
• MockAmpServer --amp-type NA500 now runs a session-based server (NA500Server). It matches the device on all 152 compared behaviors, including:
  • one client per port per session;
  • commands with no reply or a double reply;
  • leftover packets after cmd_Start;
  • front-end state;
  • the ~8000 samples/s delivery ceiling;
  • sample loss when a client stops reading.
• docs/NA500_AmpServer.md documents the protocol as observed.

App fixes (NA500)

• EEG pinned at full scale: EEG stayed at −375000 µV until the buffered reference is enabled. Every cmd_Stop turns it off, and switch commands sent within ~20 ms of power-on are lost. The app now waits 200 ms after power-on and enables the buffered reference before each start.
• Stale packets: the app discards the leftover packets the amp resends after cmd_Start. They had carried stale samples into LSL and caused a false "Packet(s) Dropped".
• Physio16 wait: the app no longer waits 5 s for a Physio16 notification the NA500 never sends. Leaving the data port unread that long made the amp drop ~4 s of samples.
• Physio streaming: the app now streams the 16 physio channels as PIB1–16. Their gains persist on the amp across sessions, so the app sets them to 12 before each start.
• Amp clock: the app sets the amp clock with cmd_SetAmpTime at power-on, as Net Station does.
• Rates: 8000 and 16000 Hz are no longer offered. The amp delivers at most ~7995 samples/s and drops the rest without any gap in the packet counter.
• System delay: systemdelayms is now optional, with a per-model default.
  • NA500: 0.8 ms + one sample period, a little under the measured ~1.0 ms + 1.1 sample periods.
  • All other models: 4.5 ms, as before.

Latency (audio tone test, relative to DAC-stamped tone markers)

Rate	EEG	Physio (PIB1)
1000 Hz	+0.15 ms (std 0.13)	+0.15 ms (std 0.13)
4000 Hz	+0.29 ms (std 0.07)	+0.21 ms (std 0.07)

There is no physio-vs-EEG skew on the NA500, and no FPGA filter delay. So decimated and native modes stream identically.

Also

• Bumps liblsl to v1.18.0.b6, which master already has via Update liblsl to v1.18.0.b6 #19.
• Documents finding the NA500 by mDNS (NA500-256ch-<serial>.local).

Triggers and impedance (NA500)

• Triggers:
  • DB9 digital inputs enabled with cmd_SetDinEnable; without it they read as all 16 active.
  • BNC in, in event mode, is reported in the _DIN stream as bit 16. Verified with a BNC out → BNC in loopback at 1 and 4 kHz.
  • The EEG trails a BNC-in event by ~1.5 samples, the ADC filter's delay. No correction is applied, as in Net Station.
• Impedance mode:
  • DAC amplitude 1023, giving ~5000 µV p-p; uses Net Station's fixed 5000 µV ideal signal.
  • The oscillator is switched on after amplitude and frequency, because the NA500 latches them at switch-on.
  • The reference is measured in the NA500's COM/REF measurement mode.
  • The acquisition state is restored explicitly afterwards.
  • Verified on a test jig that shorts all inputs: every channel ~0 kΩ, reference ~0.5 kΩ.
• Mock: models the above. A mock bug that ignored per-channel drive/10K switches is fixed; it also affected the NA400 mock.

Not covered

• Impedance accuracy: not checked against real electrodes or Net Station; only a shorting jig was available.
• DB9 pins: the physical pin-to-DIN mapping is not verified, for lack of a mating connector.
• NA400: could not be re-tested; no hardware is available.