Aravis: add the capabilities a camera is usually asked for

[HazenBabcock]

I read through the changes that Claude code made and did some manual
testing with a Basler camera and Micro-Manager. I believe that these
capabilities work as described below.

Claude code:

Follow-up to #995. Eleven commits: ten are a capability the adapter did not
have, or a bug found while adding one, and the last only moves code. Tested
against three real cameras — a Basler acA1440-220uc, an Allied Vision 1800
U-1240m and a FLIR Blackfly S BFS-U3-19S4M, all USB3 Vision — and against
emulated GigE cameras for what hardware here cannot show.

Every Bayer layout, not just RG

The layout in a Bayer format's name says which pixel of the 2x2 cell is red.
That matters to whatever demosaics the image; to this adapter, which hands the
mosaic over as it arrives, all four layouts are one component of the format's
depth. Only RG was listed, so a camera offering BayerBG8 and BayerBG12 had
both dropped from PixelFormat, and a camera with nothing else could not be
opened at all.

A frame rate the user can set

The adapter's only use of the camera's rate was to switch the limit off at
every exposure change, so that exposure paced the camera. Anyone who wanted a
fixed rate — to pace an experiment, or to stay inside a link's bandwidth — had
no way to ask for one.

AcquisitionFrameRate is now a property wherever the camera has it, with the
camera's own bounds, and AcquisitionFrameRateEnable alongside it where that
exists. SetExposure() leaves the limit alone while it is switched on, rather
than taking away a rate the user just set at the next exposure change.

Auto exposure, back again

Initialize() switches auto exposure off, because Micro-Manager is what sets
the exposure and a camera adjusting it underneath makes SetExposure() a
refusal on some cameras and a polite fiction on others. That is the right
default — but the mode was never offered afterwards, so a camera that ships in
auto, as the Blackfly S does, had it switched off at every open with no way to
ask for it back. ExposureAuto is a property now, in the camera's vocabulary
of Off, Once and Continuous. GainAuto keeps its older AUTO_* values:
changing those would break saved configurations.

Temperature, mirroring, test pattern

CCDTemperature, read-only, from the camera's DeviceTemperature. Which
sensor it reports is the camera's DeviceTemperatureSelector to say, and this
does not touch it.

ReverseX and ReverseY, where the camera has them. On a microscope the
image's handedness depends on how many mirrors the light met on the way, and a
camera that can undo that costs nothing to ask. Micro-Manager's own
TransposeMirror properties turn the image over for the screen; these turn it
over in the camera.

A test pattern, under whichever name the camera uses: the standard
TestPattern, or Basler's TestImageSelector, which is all a Basler has. The
property is TestPattern either way, carrying the entries the camera itself
offers.

GigE packet size and delay

A packet larger than the path allows is dropped by the network rather than
refused by the camera, so the stream fails on large frames and works on small
ones, with nothing in the log to say why. The delay between packets is what
keeps a camera from overrunning a switch or a host. Aravis reads both and
never negotiates them. Both appear only on a GigE camera, and only where the
camera has the feature.

What the stream did

Packet loss shows up as frames that never arrive and nothing else:
Micro-Manager simply receives fewer images than it asked for. Completed,
failed and underrun frame counts are read-only properties now. They are taken
from the stream before it is released, because the end of a sequence that went
wrong is exactly when someone goes looking for them — and a released stream
cannot be asked at all.

A Boolean that is not one

Found while testing the above. ReverseX, ReverseY and
AcquisitionFrameRateEnable are Boolean features in the standard, but a node
can carry one of those names and be something else — an integer holding 0 or

1. Asking whether the feature is available says yes; Aravis's boolean calls
   then refuse the node, and a property that reads it at every refresh turns that
   refusal into a log line per poll. One camera here produced 40 lines over 20
   property refreshes. The three properties are now offered only where the node
   really is a Boolean.

A property that describes the camera

Also found while testing. The frame rate a camera can reach moves with the
exposure, the region and the binning: an Alvium manages 15.97 Hz at a 5 ms
exposure and 5.00 Hz at 200 ms. Limits read once at startup go stale, and
Micro-Manager refuses a value outside a property's limits before the adapter
is called at all -- so a rate the camera would have accepted came back as
"Cannot set property". The limits are re-read whenever the exposure or
geometry changes, and a rate that arrives anyway is clamped to what the camera
can do now, as SetExposure() has always clamped the exposure.

The new properties also started from values written in the adapter rather than
read from the camera: a frame rate of "0.0", outside its own limits, a
temperature of 0, mirroring reported as off on a camera that was mirroring.
Micro-Manager shows a property from its cache until something refreshes it, so
that placeholder is what the user saw. Each now starts from what the camera
reports.

The last commit only moves code

AravisCamera.cpp says "These are in alphabetical order" above its method
definitions, and the commits above had put each new method next to a related
one instead. The final commit sorts them, which is about 300 lines of movement
and no change to any of them: every definition is byte for byte what it was,
and three that were already out of place before this branch are sorted too.
Nothing to read there.

Testing

Basler acA1440-220uc, colour, USB3. The frame rate holds 5 Hz across an
exposure change, with the enable flag still set afterwards. CCDTemperature
reads 61 C, matching arv-tool. ReverseX and ReverseY mirror the scene's
own pixels. The test pattern appears under Basler's TestImageSelector
spelling, offering Testimage1 to Testimage6, and frames are bit-identical with
one on where the lens sees noise with it off. The stream counters report 5
completed after the stream has been released. ExposureAuto goes Off ->
Continuous -> Off. No packet properties, because USB3 has no packets to size.
No Aravis log line for the whole run, and all 33 recorded settings as found
afterwards.

Allied Vision 1800 U-1240m, mono, USB3. The same, with the test pattern
under the standard TestPattern name instead, mirroring verified on the
scene, temperature 43 C, and 30 settings as found. One difference: this camera
refuses to let Aravis write its frame rate at all, answering USB3Vision write_memory error (write-protect). That is not the adapter -- arv-tool
cannot write it either, nor can Aravis with the enable flag set, with
AcquisitionFrameRateMode at its only value, after a delay, or with a stream
open. The property is offered, the camera declines, and the refusal is logged
once per attempt.

FLIR Blackfly S BFS-U3-19S4M, mono, USB3, and the only camera here that
ships with auto exposure and auto gain switched on -- which is what the
ExposureAuto commit is for. The frame rate holds 5 Hz across an exposure
change. The test pattern is SensorTestPattern. ReverseX and ReverseY
both mirror the scene's own pixels. With no lens fitted, where the sensor sees
a flat field, ReverseX still turns that test pattern into exactly its mirror
image byte for byte, while ReverseY cannot be judged from it -- the pattern
is vertical bars, identical flipped top to bottom. Temperature 51.5 C, no
Aravis log line, 37 settings as found.

On emulated GigE cameras, new checks covering a frame rate that survives an
exposure change, a GigE camera's packet size, and stream counters that outlive
the stream they describe. The check that a frame rate reaches the camera was
confirmed to fail when the property's write is dropped.

Each camera's settings are recorded before the test and restored afterwards,
and checked against a record of the camera as first found.

🤖 Generated with Claude Code

[marktsuchida]

Thanks, @HazenBabcock. These look like useful features.