Auto-align: level the sample on the DC bias levels from the camera window

The operator frames a good spot, confirms the two DC levels the detector
reads there, and the rig then measures its own tilt: step 1.5 mm either side
on X and then on Y, and tilt the platform until those levels come back.  The
correction that fixes an offset point is the correction that levels the whole
travel — height error and tilt effect are both proportional to the offset —
so the procedure ends by applying it and leaving it applied.

Both directions are measured from the same starting tilt and averaged, which
makes their disagreement a flatness read-out rather than something averaged
away silently.

core/auto_align.py holds the geometry and the search, Qt-free.  The three
T-axes' azimuths are the whole geometry: T1 lies along +X so it alone tilts
along X, and T0/T2 move as an equal-and-opposite pair to tilt along Y without
touching X (tilt_response derives that, and the tests pin it — an axis map
that drifts would still converge, on the wrong axis).  The search is a secant
null on the split-detector difference: probe once to learn what a microstep
is worth, sign included, then step at the null.  It refuses to servo on a
scope that has not re-triggered, escalates a probe that reads as no response
before calling an axis dead, and stops at a per-axis travel limit.

gui/align_bridge.py runs it on a worker thread; stopping is a threading.Event
rather than a queued command, because the worker is inside a long handler for
the whole run.  The camera window carries the button and the progress window,
and locks the scan panel and the jog pads while a run owns the stage.

Adds immediate MEAN measurements and an acquisition count to the scope
driver, and read_bias_mv to core/scope_inspect — the one scalar the
inspection state was missing.

KNOWN_ISSUES.md records what only the rig can settle: the probe step, the
travel limit, the hold current, and whether the piston the X phase applies
alongside its tilt matters.

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

core/scope_inspect.py

@@ -14,6 +14,11 @@ sits still.
 
 CH1 keeps the acquisition front-end so what is on screen is what a scan would
 record.  CH3 and CH4 are rescaled as DC bias monitors (see BIAS_* below).
+
+``read_bias_mv`` is the one exception to "nothing is transferred": it reads
+the two bias levels back as scalars, not waveforms, because the auto-align
+procedure (core.auto_align) has to close a loop on them.  The operator still
+watches the same screen this configures.
 """
 from __future__ import annotations
 
@@ -46,6 +51,12 @@ BIAS_POSITION_DIV = -3.5
 
 BIAS_LABELS = {3: "Bias - A", 4: "Bias - B"}
 
+# One MEAN measurement carries the shot-to-shot noise of a single record, and
+# the alignment loop has to resolve 5 mV.  The median of a handful of reads
+# rejects the odd outlier without the averaging acquisition mode, which would
+# hide exactly the intermittent response the operator is watching CH1 for.
+BIAS_READS = 5
+
 
 def inspect_channel_profiles() -> dict:
     """Channel front-end config for inspection.
@@ -102,3 +113,24 @@ def stop_inspection(scope) -> None:
     stop the sweep — it does not try to restore the acquisition profile.
     """
     scope.write("ACQuire:STATE STOP")
+
+
+def read_bias_mv(scope, reads: int = BIAS_READS) -> tuple[float, float]:
+    """Read the two DC bias levels in millivolts.
+
+    Returns ``(ch3_mv, ch4_mv)`` — DC 1 and DC 2 in the auto-align channel
+    map.  Each channel is read ``reads`` times and reduced by the median.
+
+    The two channels are read in separate batches rather than interleaved:
+    switching the immediate-measurement source costs a round trip, and these
+    are DC levels, so the few milliseconds between the batches are not a
+    source of error the way they would be for a transient.
+    """
+    if reads < 1:
+        raise ValueError("read_bias_mv needs at least one read per channel")
+
+    levels = []
+    for ch in BIAS_CHANNELS:
+        samples = sorted(scope.measure_immediate(ch, "MEAN") for _ in range(reads))
+        levels.append(samples[len(samples) // 2] * 1000.0)
+    return levels[0], levels[1]