feat: device property readout, and the display range as a histogram panel

[pskeshu]

Two commits, both from working remotely on the scope.

e2fb62b — read-only device property dump

GET /api/devices/properties?device=<label>, or no device for the list.

Gently models four properties on the Coherent Scientific Remote — the per-line PowerSetpoint (%) — because those are the four it writes. The adapter exposes many more, and among them are the ones that say whether a laser is actually emitting: enable state, interlock, faults, power readback. None of it was reachable, so #106 has been argued from what the code commands rather than what the hardware reports.

The gap matters most right now: the operator is on this microscope remotely, and cannot verify the light sheet by eye. A camera only reports photons that reach it, which needs the specimen in focus — and the light-sheet view is precisely the surface that cannot be judged out of focus. The controller can report its own state regardless.

Each property returns its value, its allowed values where the adapter declares them, and a read_only flag. Allowed values turn an opaque string into a state machine; read_only marks a readback rather than a setting, which is the distinction #106 turns on.

No require_control — it commands nothing.

4ed9c2f — display range becomes a histogram panel

Two things were wrong with the old control, and the operator named both: it sat on top of the image, and the image is the main thing anyone interacts with.

Nothing sits on the pixels now. The controls moved out of the frame into a panel attached beneath it. The overlay had also been swallowing clicks in the bottom strip of the image — a cost already conceded as a ponytail: note, which was the tell it was in the wrong place. ImageView.attach(id, { controlsHost }) takes an external host. The zoom badge stays, because it reports the view rather than controlling it.

The artefact is the status. Modelled on ImageJ/Fiji's Brightness/Contrast, at the operator's suggestion. Its insight is that the histogram and the mapping belong in one picture: the transfer line is drawn across the data it applies to, so where black and white fall relative to where the signal sits is a single glance. B&C carries no numeric status readout, and neither does this.

• Log counts — a microscopy histogram is dominated by background, and on a linear axis the part that matters is a flat line along zero.
• Auto takes the 0.1st/99.9th percentiles, not min/max: a few hot pixels must not define white.
• Computed client-side from the rendered frame via an offscreen canvas. No endpoint, no traffic, throttled to 400ms.

It is honest about what it shows: a histogram of the frame as displayed — 8-bit, already percentile-stretched per frame by the device layer. Not raw camera counts, and the panel says so. That is the argument for #149, not a limitation to hide: nuclei jammed against the top of the range become something an operator can see rather than something inferred from reading the encoder.

---

ruff, format and both mypy runs pass. JS suites 7 + 13 + 32.