NanoDrop 1000: restore macOS fixes, measure both optical paths, averaging, crash-safe teardown

[vcjdeboer]

Follow-up to #1166, from running the backend against a physical ND-1000 on macOS.

1. A regression: the macOS fixes from #1252 are not on main

#1252 was merged into add-nd1000-support before #1166 was squashed into main, but its
changes did not survive. On current main:

• _configure_usb_device calls clear_halt unguarded. On macOS libusb raises
  [Errno 2] Entity not found for an endpoint that is not actually halted, so setup()
  aborts before it can start. The backend cannot be used on macOS at all.
• stop() no longer performs the USB reset. On this firmware [0x03, 0x00] alone does
  not clear the lamp latch, so the xenon lamp can be left running.

The first commit restores both. It is self-contained and can be cherry-picked on its own
if you would rather take the fix now and review the rest separately.

2. The pedestal solenoid is a path-length selector

Released, the sample column spans the full 1 mm gap; energised, the solenoid pulls the
lever onto a fixed mechanical stop and compresses the column to 0.2 mm. There is one
stop, so exactly two paths — no argument byte to 0x0F produces a third position.

Measured on the instrument with paracetamol, in the 260–280 nm window where neither path
is saturated, the long/short absorbance ratio is 5.5, consistent with 1.0 / 0.2.
Outside that window the ratio is meaningless because one path or the other is out of
range, which is itself the reason auto-ranging exists.

Previously take_blank() and measure_absorbance() both wrapped every acquisition in
set_magnet(True), so every measurement silently used the 0.2 mm short path — the
least sensitive one — and the returned absorbance carried no record of which path
produced it, which makes it uninterpretable and incomparable.

• take_blank() blanks both paths by default. A sample measured on one path cannot be
  divided by a blank taken on another. Long path first: compressing a liquid column is
  safe, letting it expand can break it.
• measure_absorbance(path=...) takes "both" (default), "auto", "long", "short"
  or a length in mm, and returns (wavelengths, {path_mm: absorbance}). A normal run
  measures both, as the instrument does — the vendor UV-Vis screen plots them together
  from one cycle, its y-axis reading "0.1 (RED) & 1 mm Absorbance (BLACK)". Both
  readings come from a single sample loading.
• "auto" reads the long path and compresses only if it exceeds AUTORANGE_CUTOFF_AU.
  That threshold is the vendor's own: "Adjusted long path cutoff to 1.2AU for column
  formation test", ND-1000 operating software V3.5.2 release notes. The decision ignores
  the photon-starved deep UV below 235 nm, which is noise.
• select_path() picks the longest path still in range and warns when none is.
• to_path_length() gives the 10 mm cuvette equivalent, ×10 long / ×50 short.
• dark_spectrum / blank_spectrum are kept as properties onto the short-path baseline,
  which is exactly what they meant before path selection existed.

This changes the return type of measure_absorbance from a 2-tuple to
(wavelengths, {path_mm: absorbance}). Deliberate while the backend is new: an
absorbance without its path length cannot be interpreted. Happy to revisit the shape.

3. Host-side averaging

averages=N on take_blank() and measure_absorbance(). All scans in an average happen
inside one lamp-on, magnet-settled window — re-seating the solenoid between scans would
average over two different liquid columns rather than reducing read noise on one.

Measured (paracetamol, 0.2 mm, 20 ms), baseline noise 400–700 nm: averages=1 → 0.0083 A,
averages=10 → 0.0064 A. That is 1.30×, not √10 = 3.16×, and the shortfall is expected
rather than a defect: absorbance is -log10((S-D)/(B-D)), so averaging only the sample
leaves the blank's noise in place. With comparable noise the prediction is
√(2/1.1) = 1.35×. Blanking with averages=N too is needed for the full benefit, and the
docstring says so.

Smoothing is deliberately not included — Savitzky-Golay or boxcar is a presentation
choice for analysis code, not something a driver should bake into returned data.

4. async with, so a crash cannot leave the lamp burning

Every time this instrument was left with its lamp energised during development, the cause
was a script that died before reaching stop() — an unrelated AttributeError, a
failed probe. stop() itself is reliable; it just has to be reached. Reproduced on
demand, and confirmed fixed:

async with ThermoFisherNanoDrop1000() as nd:
    await nd.take_blank()
    wavelengths, spectra = await nd.measure_absorbance()

One thing I tried and deliberately left out: a process-wide failsafe on atexit. Do not
do this — an atexit handler runs during interpreter finalisation, when libusb's context
is already being torn down, and dev.write() there segfaults the process (exit code
139, with the handler's own log line printed immediately before the crash). A segfault is
not catchable, so wrapping it in try/except accomplishes nothing, and such a guard is
worse than none. A sys.excepthook version does work, but overriding a global hook as an
import side effect seemed too invasive to propose here. Happy to add it if you want it.

5. Tests

The backend had none. 12 unit tests covering the parts that are pure computation and need
no USB transport. pylabrobot/thermo_fisher/ passes: 79 passed, 4 skipped.

Verification

All of the above was exercised against a physical ND-1000 on macOS: both baselines
acquired, both paths measured from one loading, forced and auto modes, select_path
rejecting a saturated long path, teardown via __aexit__, and the crash reproduction
before and after.

ruff check and ruff format clean.

🤖 Generated with Claude Code