Updates

flicker_watch.py

@@ -24,10 +24,10 @@ import tty
 from datetime import datetime
 from pathlib import Path
 
+import numpy as np
 import requests
 import vxi11
 
-import ai_mgmt
 from csv_preprocessor import analyze_lp_file
 
 # ---------------------------------------------------------------------------
@@ -41,24 +41,45 @@ DATA_DIR    = Path(__file__).parent / "data"
 FLICKER_DIR = DATA_DIR / "flicker"
 GOOD_DIR    = DATA_DIR / "good"
 
-# LP capture parameters (matched to mipi_test_interactive.py)
+# Trigger mode:
-LP_SCALE       = 1e-6        # 1 µs/div → 20 µs window
+#   "LP_DAT"      — falling-edge on DAT0+ (CH3) crossing 0.6 V.  Fires on every
-LP_POINTS      = 200_000
+#                   LP-to-HS transition (≈ line rate, 48 kHz).  Use to sample
-LP_TRIG_OFFSET = 9e-6        # 1 µs pre / 19 µs post-trigger
+#                   normal MIPI traffic and spot per-burst anomalies.
+#   "CLK_GLITCH"  — timeout trigger on CLK+ (CH1) staying HIGH > N ms.  Fires
+#                   *only* when the clock lane goes LP for longer than expected,
+#                   i.e. an actual glitch.  Pairs with sn65_monitor.py to
+#                   capture the wire-side view of a PLL-unlock event.
+TRIGGER_MODE          = "CLK_GLITCH"   # or "LP_DAT"
+# Increased from 1 ms to 100 ms.  Earlier runs at 1 ms triggered on every
+# V-blank (≈0.5/sec on this display) — far too often to be useful.  The
+# observed PLL-unlock event from sn65_monitor is ~150 ms, so 100 ms
+# discriminates real unlocks from normal MIPI line/frame breaks.
+CLK_GLITCH_HIGH_MS    = 100.0  # CLK+ HIGH longer than this fires the trigger
+
+# Capture window
+# LP_DAT mode: 1 µs/div × 20 div = 20 µs window  (50k pts → 5 GSa/s)
+# CLK_GLITCH:  20 ms/div × 20 div = 400 ms window  (200k pts → 500 kSa/s)
+#              wide enough to bracket a 150 ms event with margin on both sides
+if TRIGGER_MODE == "CLK_GLITCH":
+    LP_SCALE       = 20e-3
+    LP_POINTS      = 200_000
+    LP_TRIG_OFFSET = 0.0          # centre the trigger so we see before+after
+    SEGMENT_COUNT  = 1            # one big window per acquire is plenty
+else:
+    LP_SCALE       = 1e-6
+    LP_POINTS      = 50_000
+    LP_TRIG_OFFSET = 9e-6
+    SEGMENT_COUNT  = 100
+
 LP_V_SCALE     = 0.2
 LP_V_OFFSET    = 0.6
 LP_TRIG_LEVEL  = 0.6
 
-# Segmented memory: capture N back-to-back LP triggers per :DIGitize, then
+CYCLE_S        = 10.0
-# dump the whole acquisition as a single H5 file.  Massively higher coverage
+# CLK_GLITCH triggers can take many seconds (or never come) — give it the full
-# than single-shot CSV captures.
+# cycle.  LP_DAT triggers fill 100 segments in well under a second.
-SEGMENT_COUNT  = 100
+TRIG_TIMEOUT_S = CYCLE_S - 0.5 if TRIGGER_MODE == "CLK_GLITCH" \
-SAVE_FORMAT    = "H5"        # Keysight native multi-segment format
+                 else max(SEGMENT_COUNT * 0.020 + 10.0, 15.0)
-
-CYCLE_S        = 10.0        # seconds video is on per cycle
-# Filling N segments takes ~N × LP-trigger period.  LP triggers fire roughly
-# at line rate (≈48 kHz) so 100 segments fill in ms, but allow margin.
-TRIG_TIMEOUT_S = max(SEGMENT_COUNT * 0.020 + 5.0, 10.0)
 
 # ---------------------------------------------------------------------------
 # Scope setup
@@ -67,11 +88,32 @@ scope = vxi11.Instrument(SCOPE_IP)
 scope.timeout = 30
 
 
+def _drain_scpi_errors(label: str = "") -> list[str]:
+    """Pop everything from the scope's error queue; return list of error strings."""
+    errs = []
+    for _ in range(20):
+        try:
+            r = scope.ask(":SYSTem:ERRor?").strip()
+        except Exception:
+            break
+        if not r or r.startswith("0,") or r.startswith("+0,") or r == "0":
+            break
+        errs.append(r)
+    if errs and label:
+        print(f"  [{label}] SCPI errors: {errs}")
+    return errs
+
+
 def setup_scope() -> None:
     """One-shot scope init — channels, math, default trigger."""
     print("CONFIGURING SCOPE...")
+    try:
+        idn = scope.ask("*IDN?").strip()
+        print(f"  IDN: {idn}")
+    except Exception as e:
+        print(f"  IDN read failed: {e}")
     cmds = [
-        "*RST", ":RUN", ":STOP",
+        "*RST", ":RUN", ":STOP", "*CLS",
         ":CHANnel1:DISPlay ON", ":CHANnel1:INPut DC50", ":CHANnel1:PROBe 19.2",
         ":CHANnel1:LABel 'CLK+'",
         ":CHANnel2:DISPlay ON", ":CHANnel2:INPut DC50", ":CHANnel2:PROBe 19.2",
@@ -88,25 +130,93 @@ def setup_scope() -> None:
     for c in cmds:
         scope.write(c)
         time.sleep(0.05)
+    _drain_scpi_errors("setup_scope")
     print("SCOPE READY.")
 
 
+def _read_ieee_block() -> bytes:
+    """
+    Read an IEEE 488.2 definite-length binary block from the scope:
+        '#' <ndigits> <length> <data> [\\n]
+    """
+    # Read header: '#' then one digit telling us how many length-digits follow.
+    head = scope.read_raw(2)
+    if not head.startswith(b"#"):
+        # Sometimes vxi11 returns a longer chunk; locate the '#'
+        idx = head.find(b"#")
+        if idx < 0:
+            extra = scope.read_raw(64)
+            head = head + extra
+            idx = head.find(b"#")
+        head = head[idx:idx + 2]
+    ndigits = int(head[1:2])
+    if ndigits == 0:
+        # "#0..." indicates indefinite-length; read until newline.
+        return scope.read_raw().rstrip(b"\r\n")
+    length_bytes = scope.read_raw(ndigits)
+    nbytes = int(length_bytes)
+    data = b""
+    while len(data) < nbytes:
+        chunk = scope.read_raw(nbytes - len(data))
+        if not chunk:
+            break
+        data += chunk
+    # Discard the trailing newline if present
+    try:
+        scope.read_raw(1)
+    except Exception:
+        pass
+    return data
+
+
 def configure_for_lp() -> None:
-    """LP-mode + segmented memory: N back-to-back LP triggers per acquisition."""
+    """LP-mode capture, with trigger configured per TRIGGER_MODE."""
     for ch in (1, 2, 3, 4):
         scope.write(f":CHANnel{ch}:SCALe {LP_V_SCALE:.3f}")
         scope.write(f":CHANnel{ch}:OFFSet {LP_V_OFFSET:.3f}")
-    scope.write(":TRIGger:EDGE:SOURce CHANnel3")
+
-    scope.write(":TRIGger:EDGE:SLOPe NEGative")
+    if TRIGGER_MODE == "CLK_GLITCH":
-    scope.write(f":TRIGger:EDGE:LEVel {LP_TRIG_LEVEL:.3f}")
+        # Pulse-width (GLITch) trigger on the Infiniium A/B (firmware 5.x):
+        # fires at the falling edge of a CH1 (CLK+) HIGH pulse longer than
+        # CLK_GLITCH_HIGH_MS — i.e. CLK held LP-11 for an unusually long time.
+        # The newer :TRIGger:TIMeout:* SCPI is rejected by this scope (-113).
+        _drain_scpi_errors()
+        scope.write(":TRIGger:MODE GLITch")
+        scope.write(":TRIGger:GLITch:SOURce CHANnel1")
+        scope.write(":TRIGger:GLITch:POLarity POSitive")
+        scope.write(":TRIGger:GLITch:DIRection GREaterthan")
+        scope.write(f":TRIGger:GLITch:WIDTh {CLK_GLITCH_HIGH_MS * 1e-3:.3E}")
+        scope.write(f":TRIGger:GLITch:LEVel CHANnel1,{LP_TRIG_LEVEL:.3f}")
+        time.sleep(0.2)
+        errs = _drain_scpi_errors()
+        if errs:
+            print(f"  GLITch trigger setup SCPI errors: {errs}")
+        try:
+            mode = scope.ask(":TRIGger:MODE?").strip()
+            w    = scope.ask(":TRIGger:GLITch:WIDTh?").strip()
+            print(f"  GLITch trigger:  mode={mode}  CLK+ HIGH > {float(w)*1000:.1f} ms")
+        except Exception as e:
+            print(f"  GLITch trigger readback failed: {e}")
+    else:
+        # Edge trigger on falling DAT0+: fires on every LP-to-HS transition.
+        scope.write(":TRIGger:MODE EDGE")
+        scope.write(":TRIGger:EDGE:SOURce CHANnel3")
+        scope.write(":TRIGger:EDGE:SLOPe NEGative")
+        scope.write(f":TRIGger:EDGE:LEVel {LP_TRIG_LEVEL:.3f}")
+
     scope.write(":TRIGger:SWEep NORMal")
     scope.write(f":TIMebase:SCALe {LP_SCALE:.3E}")
     scope.write(f":ACQuire:POINts {LP_POINTS}")
     scope.write(f":TIMebase:POSition {LP_TRIG_OFFSET:.2E}")
-    # Segmented memory: fill N segments per :DIGitize.
+
-    scope.write(":ACQuire:MODE SEGMented")
+    if SEGMENT_COUNT > 1:
-    scope.write(f":ACQuire:SEGMented:COUNt {SEGMENT_COUNT}")
+        scope.write(":ACQuire:MODE SEGMented")
+        scope.write(f":ACQuire:SEGMented:COUNt {SEGMENT_COUNT}")
+    else:
+        scope.write(":ACQuire:MODE RTIMe")
+
     time.sleep(0.5)
+    _drain_scpi_errors("configure_for_lp")
 
 
 def arm_and_wait(timeout_s: float) -> bool:
@@ -138,14 +248,69 @@ def arm_and_wait(timeout_s: float) -> bool:
             pass
 
 
-def save_lp(base_name: str) -> None:
+def _fetch_channel_segments(channel: int, n_segments: int):
-    """Save all N segments of Ch1 (CLK+) and Ch3 (DAT0+) as a single H5 each."""
+    """
-    base = f"C:\\TEMP\\{base_name}"
+    Read all segments for one channel via :WAVeform:DATA?.  Returns
-    ext  = SAVE_FORMAT.lower()
+    (times_ndarray, list_of_volts_ndarrays).  Time axis is shared across all
-    scope.write(f':DISK:SAVE:WAVeform CHANnel1,"{base}_clk.{ext}",{SAVE_FORMAT}')
+    segments.  When n_segments == 1 we skip the SEGMented:INDex select since
-    time.sleep(3.0)
+    we may be in RTIMe (single-shot) mode rather than SEGMented mode.
-    scope.write(f':DISK:SAVE:WAVeform CHANnel3,"{base}_dat.{ext}",{SAVE_FORMAT}')
+    """
-    time.sleep(3.0)
+    import numpy as np
+    scope.write(f":WAVeform:SOURce CHANnel{channel}")
+    scope.write(":WAVeform:FORMat WORD")
+    scope.write(":WAVeform:BYTeorder LSBFirst")
+
+    x_inc = float(scope.ask(":WAVeform:XINCrement?"))
+    x_org = float(scope.ask(":WAVeform:XORigin?"))
+    y_inc = float(scope.ask(":WAVeform:YINCrement?"))
+    y_org = float(scope.ask(":WAVeform:YORigin?"))
+
+    segs: list = []
+    for i in range(1, n_segments + 1):
+        if n_segments > 1:
+            scope.write(f":ACQuire:SEGMented:INDex {i}")
+        scope.write(":WAVeform:DATA?")
+        raw = _read_ieee_block()
+        codes = np.frombuffer(raw, dtype="<i2")
+        volts = codes.astype(np.float64) * y_inc + y_org
+        segs.append(volts)
+
+    n = len(segs[0]) if segs else 0
+    times = np.arange(n) * x_inc + x_org
+    return times, segs
+
+
+def save_lp(base_name: str) -> tuple[bool, list[str]]:
+    """
+    Read all N segments for CLK and DAT directly via VXI-11 binary transfer
+    and write per-segment CSVs locally to DATA_DIR.
+
+    Returns (ok, errs).  Filenames match csv_preprocessor's expected pattern:
+        {base_name}_seg{NNN}_{clk|dat}.csv
+    """
+    import numpy as np
+    _drain_scpi_errors()
+    try:
+        t_clk, clk_segs = _fetch_channel_segments(1, SEGMENT_COUNT)
+        t_dat, dat_segs = _fetch_channel_segments(3, SEGMENT_COUNT)
+    except Exception as e:
+        return (False, [f"fetch error: {e}"])
+
+    errs = _drain_scpi_errors()
+
+    n_written = 0
+    for i, (clk, dat) in enumerate(zip(clk_segs, dat_segs), start=1):
+        clk_path = DATA_DIR / f"{base_name}_seg{i:03d}_clk.csv"
+        dat_path = DATA_DIR / f"{base_name}_seg{i:03d}_dat.csv"
+        np.savetxt(clk_path, np.column_stack([t_clk, clk]),
+                   delimiter=",", fmt="%.6e")
+        np.savetxt(dat_path, np.column_stack([t_dat, dat]),
+                   delimiter=",", fmt="%.6e")
+        n_written += 1
+
+    if n_written == 0:
+        return (False, errs or ["no segments written"])
+    return (True, errs)
 
 
 # ---------------------------------------------------------------------------
@@ -186,48 +351,6 @@ def video_stop() -> None:
         print(f"  VIDEO STOP failed: {e}")
 
 
-# ---------------------------------------------------------------------------
-# H5 transfer (ai_mgmt only handles CSV — segmented mode produces .h5)
-# ---------------------------------------------------------------------------
-def _transfer_h5_files() -> int:
-    """SMB-pull every .h5 from the scope share into DATA_DIR; delete on scope."""
-    from smb.SMBConnection import SMBConnection
-    import socket
-    conn = SMBConnection(
-        ai_mgmt.USERNAME, ai_mgmt.PASSWORD,
-        socket.gethostname(), ai_mgmt.SERVER_NAME,
-        use_ntlm_v2=True, is_direct_tcp=True,
-    )
-    if not conn.connect(ai_mgmt.SERVER, 445):
-        print("  H5 transfer: could not connect to scope share")
-        return 0
-    count = 0
-    try:
-        h5_paths: list[str] = []
-        def walk(path: str) -> None:
-            for entry in conn.listPath(ai_mgmt.SHARE, path):
-                if entry.filename in (".", ".."):
-                    continue
-                full = f"{path}/{entry.filename}"
-                if entry.isDirectory:
-                    walk(full)
-                elif entry.filename.lower().endswith(".h5"):
-                    h5_paths.append(full)
-        walk("/")
-        for remote in h5_paths:
-            local = DATA_DIR / Path(remote).name
-            try:
-                with open(local, "wb") as fh:
-                    conn.retrieveFile(ai_mgmt.SHARE, remote, fh)
-                conn.deleteFiles(ai_mgmt.SHARE, remote)
-                count += 1
-            except Exception as e:
-                print(f"  H5 transfer failed for {Path(remote).name}: {e}")
-    finally:
-        conn.close()
-    return count
-
-
 # ---------------------------------------------------------------------------
 # Register snapshot from device (DSIM PHY + SN65DSI83)
 # ---------------------------------------------------------------------------
@@ -297,39 +420,14 @@ def archive_and_analyse(event: str, since_iso: str) -> None:
     # Register snapshot first (fast, before scope transfer which takes longer)
     fetch_registers_snapshot(target, event_ts)
 
-    print(f"  Transferring scope → {target} ...")
+    # Segment CSVs are already in DATA_DIR (written directly by save_lp via
-    try:
+    # SCPI binary read).  Just move the ones from this event into the folder.
-        copied, failed = ai_mgmt.transfer_csv_files()
-    except Exception as e:
-        print(f"  TRANSFER ERROR: {e}")
-        return
-    print(f"  {copied} file(s) transferred  ({failed} failed)")
-
-    # ai_mgmt only fetches CSVs.  H5 (segmented) files need a separate pass.
-    h5_count = _transfer_h5_files()
-    if h5_count:
-        print(f"  {h5_count} H5 file(s) transferred")
-
-    # Move just-arrived files (csv + h5) out of data/ (flat) into the event folder.
     moved = 0
-    for f in list(DATA_DIR.glob("*.csv")) + list(DATA_DIR.glob("*.h5")):
+    for f in DATA_DIR.glob("*.csv"):
         if f.is_file():
             shutil.move(str(f), target / f.name)
             moved += 1
-    print(f"  {moved} file(s) archived to {target.relative_to(DATA_DIR.parent)}")
+    print(f"  {moved} segment CSV(s) archived to {target.relative_to(DATA_DIR.parent)}")
-
-    # Explode each H5 into per-segment CSVs so csv_preprocessor can analyse them.
-    from explode_h5 import explode
-    h5_files = sorted(target.glob("*_lp_*.h5"))
-    seg_csv_count = 0
-    for h5 in h5_files:
-        try:
-            csvs = explode(h5)
-            seg_csv_count += len(csvs)
-        except Exception as e:
-            print(f"  EXPLODE error on {h5.name}: {e}")
-    if h5_files:
-        print(f"  exploded {len(h5_files)} H5 file(s) → {seg_csv_count} segment CSV(s)")
 
     if event != "flicker":
         return
@@ -416,9 +514,14 @@ def main() -> None:
                 cycle_end  = time.time() + CYCLE_S
 
                 video_start()
+                mode_desc = (
+                    f"CLK_GLITCH (CLK+ HIGH > {CLK_GLITCH_HIGH_MS:.1f} ms, "
+                    f"{LP_SCALE * 20 * 1000:.0f} ms window)"
+                    if TRIGGER_MODE == "CLK_GLITCH"
+                    else f"LP_DAT ({SEGMENT_COUNT} segs/acquire)"
+                )
                 print(f"\n[cycle {cycle:03d}  {cycle_ts}]  video ON  "
-                      f"({CYCLE_S:.0f}s window, {SEGMENT_COUNT} segs/acquire)",
+                      f"({CYCLE_S:.0f}s window, {mode_desc})", flush=True)
-                      flush=True)
 
                 event = None
                 last_tick = 0.0
@@ -429,16 +532,29 @@ def main() -> None:
 
                     if arm_and_wait(TRIG_TIMEOUT_S):
                         try:
-                            save_lp(base)
+                            ok, errs = save_lp(base)
-                            cycle_caps.append(base)
+                            if ok:
-                            print(f"  + acq {seq:02d}  ({SEGMENT_COUNT} segs)  "
+                                cycle_caps.append(base)
-                                  f"[{remaining():4.1f}s left]", flush=True)
+                                tag = ("CLK GLITCH" if TRIGGER_MODE == "CLK_GLITCH"
+                                       else f"{SEGMENT_COUNT} segs")
+                                print(f"  + acq {seq:02d}  ({tag})  "
+                                      f"[{remaining():4.1f}s left]",
+                                      flush=True)
+                            else:
+                                print(f"  ! acq {seq:02d} SAVE FAILED — "
+                                      f"{errs[0][:80] if errs else 'unknown'}",
+                                      flush=True)
                         except Exception as e:
                             print(f"  save error: {e}", flush=True)
                     else:
-                        # Trigger timed out — print a heartbeat at most every 2s
+                        # Trigger timed out — print a heartbeat at most every 2s.
+                        # In CLK_GLITCH mode this is the *normal* state: it just
+                        # means no glitch happened during this cycle.
                         if time.time() - last_tick > 2.0:
-                            print(f"  ... waiting for trigger  "
+                            msg = ("waiting for CLK glitch"
+                                   if TRIGGER_MODE == "CLK_GLITCH"
+                                   else "waiting for trigger")
+                            print(f"  ... {msg}  "
                                   f"[{remaining():4.1f}s left]", flush=True)
                             last_tick = time.time()
 

sn65_monitor.py

@@ -0,0 +1,277 @@
+#!/usr/bin/env python3
+"""
+sn65_monitor.py — High-rate SN65DSI83 register monitor.
+
+Continuously polls /sn65_registers at ~20 Hz, logs any register-state change
+in real time, and keeps a rolling 30 s window in memory.  When you press
+`f` (flicker) or `g` (good), the window is dumped to a JSON file and
+summarised so you can see whether anything moved at the moment of the event.
+
+This complements flicker_watch.py: run it in a second terminal during a
+test session to catch transient register changes that disappear before the
+post-event snapshot in flicker_watch can fetch them.
+
+Keys:
+    f  — flicker event:  dump rolling buffer + summary, keep monitoring
+    g  — good baseline:  dump rolling buffer + summary, keep monitoring
+    q  — quit
+"""
+
+from __future__ import annotations
+
+import json
+import select
+import sys
+import termios
+import time
+import tty
+from collections import deque
+from datetime import datetime
+from pathlib import Path
+
+import requests
+
+DEVICE_BASE = "http://192.168.45.8:5000"
+SN65_EP     = f"{DEVICE_BASE}/sn65_registers"
+DSIM_EP     = f"{DEVICE_BASE}/registers"
+DATA_DIR    = Path(__file__).parent / "data" / "sn65_log"
+
+# Aim for ~100 Hz SN65 polling — actual rate is bounded by the I2C-read
+# latency of the device server.  At 20 Hz the unlock pulse-width was
+# unresolvable ("≤ 50 ms"); at 100 Hz we should see whether it's e.g. 5 ms
+# or 30 ms, which narrows the root-cause search.
+POLL_DT_S       = 0.01     # 100 Hz target
+HISTORY_S       = 30.0
+HTTP_TIMEOUT_S  = 0.2      # tighter timeout — a slow read shouldn't stall the loop
+
+# DSIM register read goes through memtool and adds latency.  The current
+# endpoint only exposes 3 static PHY-timing config registers anyway, so
+# poll it once every N SN65 polls (set to 0 to disable entirely).  When the
+# device endpoint gains DSIM_STATUS / DSIM_CLKCTRL / DSIM_INTSRC / DSIM_FIFOCTRL,
+# raise this rate.
+DSIM_POLL_EVERY = 50       # at 100 Hz, every 50th poll → 2 Hz DSIM
+
+# csr_e5 error bit names from the device's register decode
+ERROR_BITS = ("pll_unlock", "cha_sot_bit_err", "cha_llp_err",
+              "cha_ecc_err", "cha_lp_err", "cha_crc_err")
+
+
+# ---------------------------------------------------------------------------
+# Non-blocking keyboard
+# ---------------------------------------------------------------------------
+class KeyReader:
+    def __enter__(self):
+        self.fd  = sys.stdin.fileno()
+        self.old = termios.tcgetattr(self.fd)
+        tty.setcbreak(self.fd)
+        return self
+
+    def get_key(self) -> str | None:
+        if select.select([sys.stdin], [], [], 0)[0]:
+            return sys.stdin.read(1).lower()
+        return None
+
+    def __exit__(self, *_):
+        termios.tcsetattr(self.fd, termios.TCSADRAIN, self.old)
+
+
+# ---------------------------------------------------------------------------
+# Register parsing
+# ---------------------------------------------------------------------------
+def extract_state(sn65_data: dict, dsim_data: dict | None) -> dict:
+    """Pull just the bits we care about into a hashable dict."""
+    regs   = sn65_data.get("registers", {}) if isinstance(sn65_data, dict) else {}
+    csr_0a = regs.get("csr_0a", {}) or {}
+    csr_e5 = regs.get("csr_e5", {}) or {}
+    state = {
+        "csr_0a":   csr_0a.get("value"),
+        "csr_e5":   csr_e5.get("value"),
+        "pll_lock": csr_0a.get("pll_lock"),
+        "clk_det":  csr_0a.get("clk_det"),
+    }
+    for k in ERROR_BITS:
+        state[k] = csr_e5.get(k)
+
+    # DSIM register values (whatever the endpoint exposes).  Currently:
+    # DSIM_PHYTIMING (0x32e100b4), DSIM_PHYTIMING1 (0x32e100b8), DSIM_PHYTIMING2 (0x32e100bc).
+    # These shouldn't change at runtime — but if any DOES move during an unlock
+    # event, that's a clue.  When the endpoint is extended to expose status
+    # registers (DSIM_STATUS / DSIM_CLKCTRL / DSIM_INTSRC / DSIM_FIFOCTRL),
+    # they'll be picked up here automatically.
+    if isinstance(dsim_data, dict):
+        for entry in dsim_data.get("registers", []) or []:
+            if isinstance(entry, dict) and "name" in entry and "value" in entry:
+                state[f"dsim_{entry['name']}"] = entry["value"]
+    return state
+
+
+def state_str(s: dict) -> str:
+    """Compact one-line representation of a state."""
+    pll = "PLL✓" if s.get("pll_lock") else "PLL✗"
+    clk = "CLK✓" if s.get("clk_det")  else "CLK✗"
+    errs = [k for k in ERROR_BITS if s.get(k)]
+    err_str = (",".join(errs) if errs else "no_err")
+    return (f"{pll} {clk} csr0a={s.get('csr_0a')} csr_e5={s.get('csr_e5')} "
+            f"{err_str}")
+
+
+# ---------------------------------------------------------------------------
+# Event handling
+# ---------------------------------------------------------------------------
+def save_event(event: str, history: deque, session_changes: list) -> Path:
+    DATA_DIR.mkdir(parents=True, exist_ok=True)
+    ts = datetime.now().strftime("%Y%m%d_%H%M%S")
+    out = DATA_DIR / f"{ts}_{event}.json"
+
+    snapshot = list(history)
+    payload = {
+        "event":           event,
+        "saved_at":        ts,
+        "n_samples":       len(snapshot),
+        "window_seconds":  HISTORY_S,
+        "samples":         snapshot,
+        "session_changes": session_changes[-200:],
+    }
+    out.write_text(json.dumps(payload, indent=2, default=str))
+
+    # Quick console summary
+    states_in_window = []
+    for s in snapshot:
+        if "state" in s:
+            sig = json.dumps(s["state"], sort_keys=True)
+            if not states_in_window or states_in_window[-1][1] != sig:
+                states_in_window.append((s["ts"], sig, s["state"]))
+
+    print(f"\n*** {event.upper()} EVENT @ {ts} ***")
+    print(f"  {len(snapshot)} samples saved → {out.relative_to(DATA_DIR.parent.parent)}")
+    if len(states_in_window) <= 1:
+        print(f"  register state was STABLE through the {HISTORY_S:.0f}s window")
+        if states_in_window:
+            print(f"    {state_str(states_in_window[0][2])}")
+    else:
+        print(f"  *** {len(states_in_window)} distinct register states seen in window: ***")
+        for ts_change, _, st in states_in_window:
+            t_iso = datetime.fromtimestamp(ts_change).strftime("%H:%M:%S.%f")[:-3]
+            print(f"    {t_iso}  {state_str(st)}")
+    return out
+
+
+# ---------------------------------------------------------------------------
+# Main
+# ---------------------------------------------------------------------------
+def main() -> None:
+    sess = requests.Session()
+    history: deque  = deque(maxlen=int(HISTORY_S / POLL_DT_S) + 10)
+    session_changes: list = []   # log of every state change since startup
+    last_state: dict | None = None
+    last_dsim:  dict | None = None
+    iter_count  = 0
+    poll_count  = 0
+    err_count   = 0
+    last_status = time.time()
+    started     = time.time()
+
+    print(f"SN65 + DSIM MONITOR")
+    print(f"  SN65: {SN65_EP}  (every poll)")
+    if DSIM_POLL_EVERY:
+        print(f"  DSIM: {DSIM_EP}  (every {DSIM_POLL_EVERY} polls)")
+    else:
+        print(f"  DSIM: disabled")
+    print(f"poll target {1.0/POLL_DT_S:.0f} Hz, rolling buffer {HISTORY_S:.0f}s")
+    print("keys: f=flicker  g=good  q=quit\n", flush=True)
+
+    with KeyReader() as keys:
+        try:
+            while True:
+                t0 = time.time()
+                iter_count += 1
+                sn65_data: dict = {}
+                err_this_poll = False
+                try:
+                    r = sess.get(SN65_EP, timeout=HTTP_TIMEOUT_S)
+                    r.raise_for_status()
+                    sn65_data = r.json()
+                except requests.exceptions.RequestException as e:
+                    err_this_poll = True
+                    history.append({"ts": t0, "error": f"sn65: {e}"})
+
+                # DSIM is fetched only every Nth iteration to keep the SN65
+                # poll rate high.  In between, we reuse the previous DSIM
+                # snapshot.
+                if DSIM_POLL_EVERY and (iter_count % DSIM_POLL_EVERY == 0):
+                    try:
+                        r = sess.get(DSIM_EP, timeout=HTTP_TIMEOUT_S)
+                        r.raise_for_status()
+                        last_dsim = r.json()
+                    except requests.exceptions.RequestException:
+                        # best-effort; keep last known
+                        pass
+                dsim_data = last_dsim
+
+                if err_this_poll:
+                    err_count += 1
+                else:
+                    state = extract_state(sn65_data, dsim_data)
+                    history.append({"ts": t0, "state": state,
+                                    "sn65_raw": sn65_data,
+                                    "dsim_raw": dsim_data})
+                    poll_count += 1
+
+                    if last_state is not None and state != last_state:
+                        delta = {k: (last_state.get(k), state.get(k))
+                                 for k in state if state.get(k) != last_state.get(k)}
+                        ts_iso = datetime.fromtimestamp(t0).strftime("%H:%M:%S.%f")[:-3]
+                        print(f"\n[{ts_iso}] CHANGE: {state_str(state)}")
+                        for k, (old, new) in delta.items():
+                            print(f"               {k}: {old} → {new}")
+                        session_changes.append(
+                            {"ts": t0, "iso": ts_iso, "delta": delta,
+                             "new_state": state}
+                        )
+                    last_state = state
+
+                # Status line every 2 s — overwrites itself with \r
+                if t0 - last_status > 2.0:
+                    rate    = poll_count / (t0 - last_status) if t0 > last_status else 0
+                    err_pct = err_count / max(1, poll_count + err_count) * 100
+                    cur     = state_str(last_state) if last_state else "(no data)"
+                    sys.stdout.write(
+                        f"\r  {rate:5.1f} Hz | err {err_pct:4.1f}% | "
+                        f"buf {len(history)} | changes {len(session_changes)} | "
+                        f"{cur}              "
+                    )
+                    sys.stdout.flush()
+                    last_status = t0
+                    poll_count  = 0
+                    err_count   = 0
+
+                # Keypress
+                key = keys.get_key()
+                if key == "f":
+                    save_event("flicker", history, session_changes)
+                elif key == "g":
+                    save_event("good", history, session_changes)
+                elif key == "q":
+                    print("\nQUIT.")
+                    break
+
+                # Pace
+                elapsed = time.time() - t0
+                if elapsed < POLL_DT_S:
+                    time.sleep(POLL_DT_S - elapsed)
+
+        except KeyboardInterrupt:
+            print("\nInterrupted (Ctrl+C).")
+
+    # Session summary
+    dur = time.time() - started
+    print(f"\n--- session summary: {dur:.1f}s, "
+          f"{len(session_changes)} state change(s) ---")
+    if session_changes:
+        print("  recent changes:")
+        for c in session_changes[-10:]:
+            print(f"    {c['iso']}  {state_str(c['new_state'])}")
+
+
+if __name__ == "__main__":
+    main()

video_cycler.py

@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+"""
+video_cycler.py — Toggle /video start/stop on the device every CYCLE_S seconds.
+
+Pairs with sn65_monitor.py: this script provokes the flicker by cycling the
+static-pink video stream, while sn65_monitor measures.  Press Ctrl+C to stop.
+"""
+
+import signal
+import sys
+import time
+from datetime import datetime
+
+import requests
+
+DEVICE_BASE = "http://192.168.45.8:5000"
+VIDEO_URL   = f"{DEVICE_BASE}/video"
+CYCLE_S     = 10.0
+HTTP_TIMEOUT_S = 3.0
+
+
+def video_start() -> None:
+    try:
+        requests.put(VIDEO_URL,
+                     json={"action": "start", "mode": "static-pink"},
+                     timeout=HTTP_TIMEOUT_S)
+    except requests.exceptions.RequestException as e:
+        print(f"  video START failed: {e}")
+
+
+def video_stop() -> None:
+    try:
+        requests.put(VIDEO_URL, json={"action": "stop"},
+                     timeout=HTTP_TIMEOUT_S)
+    except requests.exceptions.RequestException as e:
+        print(f"  video STOP failed: {e}")
+
+
+def main() -> None:
+    # On Ctrl+C, make sure we leave video stopped.
+    def _shutdown(*_):
+        print("\nshutting down — video off")
+        video_stop()
+        sys.exit(0)
+    signal.signal(signal.SIGINT,  _shutdown)
+    signal.signal(signal.SIGTERM, _shutdown)
+
+    print(f"VIDEO CYCLER — {CYCLE_S:.0f} s on / 0.5 s off (Ctrl+C to stop)\n")
+    cycle = 0
+    while True:
+        cycle += 1
+        ts = datetime.now().strftime("%H:%M:%S")
+        print(f"[{ts}] cycle {cycle:04d}  video ON", flush=True)
+        video_start()
+        time.sleep(CYCLE_S)
+        video_stop()
+        time.sleep(0.5)
+
+
+if __name__ == "__main__":
+    main()