stm32_link: sync iin_avg_ma comment with live firmware protocol

Firmware now runs BOTH iin limits (IIN_MAX reverse-feed and the IIN_MIN
forward OCP, raised to -47A) on the 8-sample boxcar; wire format is
unchanged (verified field-by-field against the live code64 console,
including a synthetic-frame decode). Note the field is a raw int16 mA
cast and wraps at +-32.768A (observed +26.5A at a true -39.0A) -- the
float iin field is the honest monitor above that.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

README.md

@@ -115,7 +115,17 @@ The GUI provides:
   protection shut the output down), the sweep aborts immediately, reports
   which protection fired (OVP / OV / OC / OP / OT), and saves the points
   collected so far instead of logging garbage rows for the rest of the
-  grid.
+  grid. A converter-output collapse (load voltage below half the expected
+  Vout while the supply is still up -- the DUT shut itself down) aborts
+  the same way; when the STM32 link is up the abort message names the
+  firmware limit/fault bits from the last frame before the brownout
+  blackout.
+- Sweep finish estimate: the status line shows `ETA hh:mm (N pts, ~M min
+  left)` next to each measured point. The remaining grid is deterministic
+  (the steps that pass the feasibility gate), but the cost per point is
+  not (settle time + instrument round-trips that vary per setup), so the
+  per-point time is measured as a running average -- thermal-hold pauses
+  excluded -- and extrapolated over the remaining feasible points.
 - Load range pinning: a mid-sweep auto-range transition on the Prodigit
   momentarily unloads the converter, so at sweep start the CC range is
   pinned to Range II for the whole run (auto-ranging restored after, with
@@ -199,9 +209,18 @@ uv run bench-plot data_20260703_140000.csv
 uv run bench-plot run1_telem.csv run2_telem.csv --vin-bin 2 --save eff.png
 
 # options: --current iout|iin, --source auto|hioki|instr|stm (data logs),
-#          --min-pout W (default 5), --vin-bin V (default 1), --save PNG
+#          --min-pout W (default 5), --vin-bin V (default 1), --save PNG,
+#          --pick / --no-pick (per-file Vin-bin checkbox picker)
 ```
 
+Multiple files are pooled into one dataset. When two files cover the same
+Vin bin, a checkbox picker opens automatically -- one row per Vin bin, one
+column per file, cell numbers showing points per bin, overlapping bins
+highlighted -- so you decide per bin which file's data to use (e.g. keep the
+re-measured 72 V column from the new run, the rest from the old one).
+`--pick` forces the picker even without overlap (useful to cut bad columns
+from a single file); `--no-pick` skips it for scripted/batch use.
+
 Also reachable via the GUI's "Plot Eff..." button (Logging section, opens the
 same dialog preselecting the last log) and `plot_eff.bat` one level up
 (double-click for the dialog, or drag && drop CSV files onto it).

testbench/gui.py

@@ -29,6 +29,7 @@ from testbench.bench import MPPTTestbench
 from testbench.gui_workers import InstrumentWorker, Cmd, STM32Worker
 from testbench.stm32_link import (
     FLAG_NAMES, FLAG_INFO_MASK, build_ping, build_clear_flags,
+    flags_to_names,
 )
 
 
@@ -156,6 +157,7 @@ class TestbenchGUI(tk.Tk):
 
         self.bench: MPPTTestbench | None = None
         self.worker: InstrumentWorker | None = None
+        self._teardown_thread: threading.Thread | None = None
         self._log_file = None
         self._log_writer = None
         self._log_count = 0
@@ -1141,6 +1143,10 @@ class TestbenchGUI(tk.Tk):
     # ── Connection ────────────────────────────────────────────────────
 
     def _connect(self) -> None:
+        if self._teardown_thread and self._teardown_thread.is_alive():
+            self._console("Previous disconnect is still releasing the "
+                          "instruments - try again in a few seconds", "warn")
+            return
         try:
             supply_addr = self._supply_addr.get().strip()
             if supply_addr == "auto":
@@ -1193,16 +1199,9 @@ class TestbenchGUI(tk.Tk):
 
     def _disconnect(self) -> None:
         self._stop_log()
-        if self.worker:
-            self.worker.stop()
-            self.worker.join(timeout=5)
-            self.worker = None
-        if self.bench:
-            try:
-                self.bench.close()
-            except Exception:
-                pass
-            self.bench = None
+        worker, bench = self.worker, self.bench
+        self.worker = None
+        self.bench = None   # _poll chain sees both None and ends
 
         self._btn_connect.config(state=tk.NORMAL)
         self._btn_setup.config(state=tk.DISABLED)
@@ -1212,7 +1211,40 @@ class TestbenchGUI(tk.Tk):
         self._load_baud.config(state=tk.NORMAL)
         self._meter_addr.config(state=tk.NORMAL)
         self._status_label.config(text="Disconnected")
-        self._console("Disconnected")
+
+        if not worker and not bench:
+            self._console("Disconnected")
+            return
+
+        # Instrument teardown runs OFF the GUI thread: with dead/absent
+        # devices every VISA/serial call blocks in multi-second timeouts,
+        # which froze the GUI ("not responding" -> apparent crash) when
+        # this was inline (2026-07-06 15:52).  Closing a VISA session
+        # while the worker is still inside a read can also crash NI-VISA
+        # natively, so if the worker won't exit the sessions are left
+        # open -- the OS reclaims them at process exit.
+        def _teardown():
+            def say(msg, tag=None):
+                try:
+                    self._console(msg, tag) if tag else self._console(msg)
+                except Exception:
+                    pass    # window already destroyed
+            if worker:
+                worker.stop()
+                worker.join(timeout=15)
+            if bench:
+                if worker and worker.is_alive():
+                    say("Instrument worker stuck in an I/O timeout - "
+                        "connections left open (freed on exit)", "warn")
+                else:
+                    try:
+                        bench.close()
+                    except Exception:
+                        pass
+            say("Disconnected")
+
+        self._teardown_thread = threading.Thread(target=_teardown, daemon=True)
+        self._teardown_thread.start()
 
     def _setup_all(self) -> None:
         self._send(Cmd.SETUP_ALL)
@@ -1715,6 +1747,25 @@ class TestbenchGUI(tk.Tk):
             return None
         return b.etemp, b.btemp
 
+    def _stm32_fault_note(self) -> str:
+        """Last STM32 fault flags as a console suffix, '' when unlinked.
+
+        On a converter collapse the MCU (fed from the Vout rail) browns
+        out, so the LAST frame before the blackout usually carries the
+        limit/fault bit that fired -- report it even when stale.
+        """
+        w = self.stm32
+        if not w:
+            return ""
+        b, wall = w.get_latest()
+        if b is None:
+            return ""
+        names = flags_to_names(b.status_flags & ~FLAG_INFO_MASK)
+        txt = ", ".join(names) if names else "no fault bits"
+        age = time.time() - wall
+        stale = f", {age:.0f}s stale" if age > TELEM_STALE_S else ""
+        return f" [STM32: {txt}{stale}]"
+
     @staticmethod
     def _supply_trip_cause(bench) -> str:
         """Best-effort query of which supply protection fired."""
@@ -1880,6 +1931,43 @@ class TestbenchGUI(tk.Tk):
             self._console(f"Load range readback failed ({e}) - range check "
                           f"skipped", "warn")
 
+        # Finish-time prediction.  The grid is deterministic (steps that
+        # pass the feasibility gate), but the cost per point is NOT: it is
+        # settle (known) + instrument round-trips (HIOKI auto-range wait +
+        # VISA latency, varies per instrument setup) + occasional holds.
+        # So the per-point time is MEASURED as a running average and
+        # extrapolated over the remaining feasible grid points.
+        def _in_range(x, stop_, step_):
+            return (x <= stop_ + step_ / 2 if step_ > 0
+                    else x >= stop_ + step_ / 2)
+
+        def _gate_ok(ll_, v_):
+            # mirror of the in-loop rejection gate (reads the live
+            # vout_est / range_max / psu_imax)
+            if range_max is not None and ll_ > range_max * 1.001:
+                return False
+            return _est_input_current(load_mode, ll_, v_, vout_est) <= psu_imax
+
+        def _pts_remaining(v_now, ll_next):
+            """Feasible grid points from (v_now, ll_next) to the end."""
+            if v_step == 0 or l_step == 0:
+                return 0
+            cnt = 0
+            vv, ll_ = v_now, ll_next
+            while _in_range(vv, v_stop, v_step):
+                while _in_range(ll_, l_stop, l_step):
+                    if _gate_ok(ll_, vv):
+                        cnt += 1
+                    ll_ += l_step
+                vv += v_step
+                ll_ = l_start
+            return cnt
+
+        self._console(
+            f"Sweep grid: {_pts_remaining(v_start, l_start)} feasible "
+            f"point(s) planned - finish estimate appears after the first "
+            f"point")
+
         bench.supply.set_current(current_limit)
         bench.supply.output_on()
         bench._apply_load_value(load_mode, l_start)
@@ -1889,6 +1977,8 @@ class TestbenchGUI(tk.Tk):
         n = 0
         v = v_start
         applied = l_start   # last load setpoint actually commanded
+        t_pt = None                  # running avg seconds per accepted point
+        t_last = time.monotonic()
 
         try:
             while not stop.is_set():
@@ -1929,9 +2019,11 @@ class TestbenchGUI(tk.Tk):
                         continue
 
                     # Pause near the firmware thermal trips (holds at ~1 A)
+                    t0_hold = time.monotonic()
                     held = self._thermal_hold(bench, load_mode, vout_est, stop)
                     if held is not None:
                         applied = held
+                        t_last += time.monotonic() - t0_hold  # ETA: skip hold
                         if stop.is_set():
                             break
 
@@ -1959,6 +2051,23 @@ class TestbenchGUI(tk.Tk):
                             f"{n} collected point(s)", "error")
                         stop.set()
                         break
+                    # Converter-output collapse abort: the DUT shut itself
+                    # down (firmware limit / fault) while the supply stayed
+                    # up -- the supply-side check above never sees this
+                    # (run 2026-07-07: Vout 48->3V at Vin=55V and the sweep
+                    # kept stepping the rest of the grid).
+                    if point.load_voltage < vout_est * 0.5:
+                        results.pop()   # this point is garbage
+                        n -= 1
+                        self._console(
+                            f"Converter output collapsed at V={v:g}V "
+                            f"{load_mode}={ll:g}{unit}: Vout "
+                            f"{point.load_voltage:.1f}V, expected ~"
+                            f"{vout_est:.0f}V{self._stm32_fault_note()} - "
+                            f"aborting sweep, saving {n} collected "
+                            f"point(s)", "error")
+                        stop.set()
+                        break
                     if point.load_voltage > 5.0:
                         vout_est = point.load_voltage
                     # Measured backstop: the estimate can be off (eff, Vout)
@@ -1980,6 +2089,17 @@ class TestbenchGUI(tk.Tk):
                         ll += l_step
                         continue
 
+                    # Update the finish-time estimate from this point's
+                    # measured wall time (thermal holds already excluded)
+                    now_m = time.monotonic()
+                    dt_pt = now_m - t_last
+                    t_last = now_m
+                    t_pt = (dt_pt if t_pt is None
+                            else 0.7 * t_pt + 0.3 * dt_pt)
+                    rem = _pts_remaining(v, ll + l_step)
+                    eta = time.strftime(
+                        "%H:%M", time.localtime(time.time() + rem * t_pt))
+
                     # Push data for live graph/readout updates
                     gui_data = {
                         "supply_V": point.supply_voltage,
@@ -2009,11 +2129,14 @@ class TestbenchGUI(tk.Tk):
                             pass
                         self._sweep_data_queue.put_nowait(gui_data)
 
+                    mins = rem * t_pt / 60.0
                     self.after(
                         0,
-                        lambda v=v, ll=ll, pt=point, n=n: self._svi_status.config(
+                        lambda v=v, ll=ll, pt=point, n=n, eta=eta, rem=rem,
+                               mins=mins: self._svi_status.config(
                             text=f"[{n}] V={v:.1f}V {load_mode}={ll:.1f}{unit} "
-                                 f"EFF={pt.efficiency:.1f}%"
+                                 f"EFF={pt.efficiency:.1f}%  ETA {eta} "
+                                 f"({rem} pts, ~{mins:.0f} min left)"
                         ),
                     )
 

testbench/plot_eff.py

@@ -8,9 +8,15 @@ Auto-detects the three CSV formats produced by the tooling:
 Left panel: operating-point scatter (x = Vin, y = current, color = efficiency).
 Right panel: efficiency vs current, one curve per Vin bin.
 
+Multiple files are pooled into one dataset. When several files contain the
+same Vin bin, a checkbox picker opens (one row per bin, one column per file)
+so you can choose which file's data to use per bin; force it with --pick or
+suppress it with --no-pick.
+
 Usage:
     bench-plot data_20260703_120000.csv
     bench-plot run_telem.csv another_telem.csv --current iin --save eff.png
+    bench-plot sweep_a.csv sweep_b.csv --pick
 """
 
 from __future__ import annotations
@@ -18,7 +24,9 @@ from __future__ import annotations
 import argparse
 import csv
 import math
+import os
 import sys
+from collections import Counter
 
 import matplotlib.pyplot as plt
 import numpy as np
@@ -45,9 +53,29 @@ def _detect_format(header: list[str]) -> str:
         return "telem"
     if "voltage_set" in cols and "efficiency" in cols:
         return "sweep"
+    if not header:
+        raise ValueError("empty file (a log that is still being written?)")
     raise ValueError(f"unrecognized CSV header: {header[:6]}...")
 
 
+def _fatal(msg: str) -> None:
+    """Print the error AND show it in a messagebox: when spawned from the
+    GUI button nobody sees stderr, so exiting silently looks like a no-op."""
+    print(msg, file=sys.stderr)
+    if not os.environ.get("BENCH_PLOT_HEADLESS"):
+        try:
+            import tkinter as tk
+            from tkinter import messagebox
+            r = tk.Tk()
+            r.withdraw()
+            r.attributes("-topmost", True)
+            messagebox.showerror("bench-plot", msg, parent=r)
+            r.destroy()
+        except Exception:
+            pass
+    sys.exit(1)
+
+
 def _extract(path: str, source: str) -> tuple[dict, str]:
     """Read one CSV -> dict of float lists (vin_V, iin_A, iout_A, eff_pct,
     p_out_W) plus the efficiency-source label actually used."""
@@ -120,6 +148,110 @@ def _extract(path: str, source: str) -> tuple[dict, str]:
     return out, labels[source]
 
 
+class _BinPicker:
+    """Checkbox grid: one row per Vin bin, one column per file.
+
+    result is a list of selected-center sets (aligned with bin_table)
+    after OK, or None if the window was cancelled/closed.
+    """
+
+    ROW_H = 26
+
+    def __init__(self, bin_table: list[tuple[str, dict]], w: float):
+        import tkinter as tk
+        from tkinter import ttk
+        self._tk = tk
+        self.root = tk.Tk()
+        self.root.title(f"Select Vin bins per file ({w:g} V bins)")
+        # Spawned from the GUI's Plot Eff button, Windows denies foreground
+        # to the new process (the user is mid-click in the GUI) and the
+        # picker opens hidden BEHIND it -- keep this short-lived dialog
+        # on top and grab focus.
+        self.root.lift()
+        self.root.attributes("-topmost", True)
+        self.root.focus_force()
+        self._n_files = len(bin_table)
+        self._vars: dict[tuple[int, float], object] = {}
+        self.result: list[set] | None = None
+
+        centers = sorted({c for _, bins in bin_table for c in bins})
+        shared = {c for c in centers
+                  if sum(c in bins for _, bins in bin_table) > 1}
+
+        top = ttk.Frame(self.root, padding=6)
+        top.pack(fill="both", expand=True)
+        canvas = tk.Canvas(
+            top, highlightthickness=0,
+            height=min(self.ROW_H * (len(centers) + 2) + 8, 560))
+        vsb = ttk.Scrollbar(top, orient="vertical", command=canvas.yview)
+        grid = ttk.Frame(canvas)
+        grid.bind("<Configure>",
+                  lambda e: canvas.configure(scrollregion=canvas.bbox("all"),
+                                             width=grid.winfo_reqwidth()))
+        canvas.create_window((0, 0), window=grid, anchor="nw")
+        canvas.configure(yscrollcommand=vsb.set)
+        canvas.pack(side="left", fill="both", expand=True)
+        vsb.pack(side="right", fill="y")
+        canvas.bind_all("<MouseWheel>", lambda e: canvas.yview_scroll(
+            -1 if e.delta > 0 else 1, "units"))
+
+        ttk.Label(grid, text="Vin bin").grid(row=0, column=0, padx=6, sticky="w")
+        for fi, (name, _bins) in enumerate(bin_table):
+            short = name if len(name) <= 26 else "..." + name[-23:]
+            ttk.Label(grid, text=short).grid(row=0, column=1 + fi, padx=8)
+            ttk.Button(grid, text="all/none", width=8,
+                       command=lambda fi=fi: self._toggle_col(fi)
+                       ).grid(row=1, column=1 + fi, padx=8)
+        for ri, c in enumerate(centers):
+            lbl = ttk.Label(grid, text=f"{c:g} V")
+            if c in shared:
+                lbl.configure(foreground="#cc6600")
+            lbl.grid(row=2 + ri, column=0, padx=6, sticky="w")
+            for fi, (_name, bins) in enumerate(bin_table):
+                if c not in bins:
+                    continue
+                var = tk.BooleanVar(self.root, value=True)
+                self._vars[(fi, c)] = var
+                ttk.Checkbutton(grid, text=str(bins[c]), variable=var
+                                ).grid(row=2 + ri, column=1 + fi,
+                                       padx=8, sticky="w")
+
+        bar = ttk.Frame(self.root, padding=(6, 0, 6, 6))
+        bar.pack(fill="x")
+        ttk.Label(bar, text="numbers = points per bin, "
+                            "orange = bin present in several files"
+                  ).pack(side="left")
+        ttk.Button(bar, text="Cancel",
+                   command=self.root.destroy).pack(side="right", padx=4)
+        ttk.Button(bar, text="Plot", command=self._ok).pack(side="right")
+
+    def _toggle_col(self, fi: int) -> None:
+        cells = [v for (f, _c), v in self._vars.items() if f == fi]
+        state = not all(v.get() for v in cells)
+        for v in cells:
+            v.set(state)
+
+    def _ok(self) -> None:
+        self.result = [
+            {c for (f, c), v in self._vars.items() if f == fi and v.get()}
+            for fi in range(self._n_files)]
+        self.root.destroy()
+
+    def run(self) -> list[set] | None:
+        self.root.mainloop()
+        return self.result
+
+
+def _pick_bins(bin_table: list[tuple[str, dict]], w: float) -> list[set] | None:
+    """Show the picker; on a headless/Tk failure fall back to everything."""
+    try:
+        picker = _BinPicker(bin_table, w)
+    except Exception as e:
+        print(f"bin picker unavailable ({e}) - using all bins")
+        return [set(bins) for _, bins in bin_table]
+    return picker.run()
+
+
 def main() -> None:
     ap = argparse.ArgumentParser(
         description="Plot efficiency vs input voltage vs current from bench CSVs.")
@@ -137,6 +269,12 @@ def main() -> None:
                     help="drop points below this output power (default 5.0)")
     ap.add_argument("--save", metavar="PNG", help="write the figure instead of showing it")
     ap.add_argument("--title", default=None, help="figure title override")
+    g = ap.add_mutually_exclusive_group()
+    g.add_argument("--pick", action="store_true",
+                   help="always show the per-file Vin-bin checkbox picker")
+    g.add_argument("--no-pick", action="store_true",
+                   help="never show the picker (default: it opens when "
+                        "several files contain the same Vin bin)")
     args = ap.parse_args()
 
     if not args.csv:
@@ -144,33 +282,65 @@ def main() -> None:
         from tkinter import filedialog
         root = tk.Tk()
         root.withdraw()
+        root.attributes("-topmost", True)
         args.csv = list(filedialog.askopenfilenames(
+            parent=root,
             title="Select logged CSV(s) to plot",
             filetypes=[("CSV files", "*.csv"), ("All files", "*.*")]))
         root.destroy()
         if not args.csv:
             sys.exit("no file selected")
 
-    data: dict[str, list] = {k: [] for k in ("vin_V", "iin_A", "iout_A", "eff_pct", "p_out_W")}
-    labels = set()
+    # Load each file separately (kept apart for the per-file bin picker)
+    w = args.vin_bin
+    parts, labels, n_total, bad = [], set(), 0, []
     for path in args.csv:
-        part, label = _extract(path, args.source)
-        for k in data:
-            data[k].extend(part[k])
+        try:
+            part, label = _extract(path, args.source)
+        except Exception as e:
+            bad.append(f"{os.path.basename(path)}: {e}")
+            print(f"  skipping {os.path.basename(path)}: {e}")
+            continue
         labels.add(label)
+        fvin = np.asarray(part["vin_V"])
+        fcur = np.asarray(part["iin_A" if args.current == "iin" else "iout_A"])
+        feff = np.asarray(part["eff_pct"])
+        fpout = np.asarray(part["p_out_W"])
+        n_total += len(fvin)
+        keep = (np.isfinite(fvin) & np.isfinite(fcur) & np.isfinite(feff)
+                & (feff > 0.0) & (feff <= 105.0) & (fpout >= args.min_pout))
+        p = {"name": os.path.basename(path), "vin": fvin[keep],
+             "cur": fcur[keep], "eff": feff[keep]}
+        p["bins"] = np.round(p["vin"] / w) * w
+        parts.append(p)
+        if len(args.csv) > 1:
+            rng = (f"Vin bins {p['bins'].min():g}..{p['bins'].max():g} V"
+                   if len(p["vin"]) else "no usable points")
+            print(f"  {p['name']}: {len(p['vin'])} pts, {rng}")
 
-    vin = np.asarray(data["vin_V"])
-    cur = np.asarray(data["iin_A" if args.current == "iin" else "iout_A"])
-    eff = np.asarray(data["eff_pct"])
-    pout = np.asarray(data["p_out_W"])
+    if not parts:
+        _fatal("none of the selected files were readable:\n" + "\n".join(bad))
 
-    keep = (np.isfinite(vin) & np.isfinite(cur) & np.isfinite(eff)
-            & (eff > 0.0) & (eff <= 105.0) & (pout >= args.min_pout))
-    n_total = len(vin)
-    vin, cur, eff = vin[keep], cur[keep], eff[keep]
+    # Per-file Vin-bin selection: opens automatically when files overlap
+    bin_table = [(p["name"], Counter(p["bins"].tolist())) for p in parts]
+    files_per_bin = Counter(c for _, bins in bin_table for c in bins)
+    overlap = any(n > 1 for n in files_per_bin.values())
+    if not args.no_pick and (args.pick or (overlap and len(parts) > 1)):
+        sels = _pick_bins(bin_table, w)
+        if sels is None:
+            sys.exit("cancelled")
+        for p, sel in zip(parts, sels):
+            m = (np.isin(p["bins"], sorted(sel)) if sel
+                 else np.zeros(len(p["bins"]), dtype=bool))
+            for k in ("vin", "cur", "eff", "bins"):
+                p[k] = p[k][m]
+
+    vin = np.concatenate([p["vin"] for p in parts])
+    cur = np.concatenate([p["cur"] for p in parts])
+    eff = np.concatenate([p["eff"] for p in parts])
     if len(vin) == 0:
-        sys.exit(f"no usable points ({n_total} rows read; all filtered — "
-                 f"check --min-pout / --source)")
+        _fatal(f"no usable points ({n_total} rows read; all filtered or "
+               f"deselected — check --min-pout / --source / bin picker)")
 
     ipk = int(np.argmax(eff))
     print(f"{len(vin)} points ({n_total - len(vin)} filtered) | "
@@ -195,7 +365,6 @@ def main() -> None:
     ax1.grid(alpha=0.3)
 
     # Right: efficiency vs current, one mean curve per Vin bin
-    w = args.vin_bin
     centers = np.unique(np.round(vin / w) * w)
     cmap = plt.cm.plasma(np.linspace(0.0, 0.9, len(centers)))
     for color, c0 in zip(cmap, centers):

testbench/stm32_link.py

@@ -164,7 +164,7 @@ class BroadcastData:
     param_value: int = 0   # raw bits; interpret via decode_param_bits(param_type, ...)
     pong: int = 0          # increments when the MCU processes CMD_PING
     sys_current_ma: int = 0   # Vout-rail housekeeping current, mA
-    iin_avg_ma: int = 0       # 8-sample boxcar of iin, mA (the IIN_MAX trip quantity)
+    iin_avg_ma: int = 0       # 8-sample boxcar of iin, mA (the IIN_MAX/IIN_MIN trip quantity)
     timestamp: float = field(default_factory=time.time)
 
     @property