fix: enable sigfm for goodix511 and fix Debian build

[Mounstroya]

Problem

On Debian with a Goodix TLS Fingerprint Sensor 511 (27c6:5110), enrollment completes successfully but fprintd-verify always returns verify-no-match.

Root cause

fpi_device_goodixtls511_class_init never sets img_dev_class->algorithm, so fp-image-device.c falls back to FPI_PRINT_NBIS (the default, value 0). The sigfm library is compiled and linked but never invoked. NBIS performs poorly on the sensor's low-resolution 80×64 images, hence the consistent match failure.

Changes

libfprint/drivers/goodixtls/goodix511.c

• Set img_dev_class->algorithm = FPI_DEVICE_ALGO_SIGFM in class init so sigfm is actually used
• Add per-frame contrast normalization in crop_frame to stretch pixel values to the full 0–255 range, improving keypoint detection on the low-contrast sensor images

meson.build

• Replace dependency('udev') with dependency('libudev') in both occurrences — on Debian the pkg-config name is libudev, not udev, causing the build to fail before reaching compilation

libfprint/sigfm/meson.build

• Mark doctest as required: false and guard the test executable behind if doctest.found() — the test framework is not needed to build or use the library

Testing

Tested on Debian 13 (trixie) with:

• Goodix TLS Fingerprint Sensor 511 (27c6:5110)
• libopencv-dev 4.10
• fprintd 1.94.5

After applying these changes, fprintd-enroll and fprintd-verify both work correctly.

[finikorg]

Independent confirmation of the root cause, on different hardware and a
different distro.

I hit the same symptom on a Huawei MateBook D (27c6:5110), Ubuntu 24.04:
enrolment reporting 20/20 while verify never matched. I arrived at the same
line you did before finding this PR — fpi_device_goodixtls511_class_init
never sets img_dev_class->algorithm, so fp-image-device.c:194 leaves it at
the FPI_PRINT_NBIS default and sigfm is linked but never called.

To be clear about what I tested: I applied only the algorithm line, not
your contrast-normalisation hunk, so the numbers below are for that change
alone. I have not evaluated the crop_frame change or the Debian build fixes.

What made it unambiguous was that fpi_print_sigfm_match() logs
fp_dbg("sigfm score %d/%d", ...) on every comparison, and with
G_MESSAGES_DEBUG=all no such line ever appeared — the function was not
being reached at all. Meanwhile the stored template serialised as
(a(aiaiai)), the NBIS format from fp-print.c:689, rather than sigfm's
(a(ay)) at fp-print.c:723, and it was 518 bytes.

Before and after, same finger, same session:

	NBIS (default)	with algorithm = SIGFM
enrolment yield	20/44, 12/66, 2/44 presses	20/20, 100%
No minutiae found errors	42 of 44 presses	0
verification	0 of 10	3 of 3
stored template	518 B, (a(aiaiai))	608 KB, (a(ay))

Scores once sigfm is actually invoked, against the enrolled samples:

sigfm score 989/24     <- match
sigfm score 441/24
sigfm score 293/24
sigfm score 44/24
sigfm score 28/24
sigfm score 0/24  (most other samples, as expected)

Worth noting for anyone tempted to lower it: bz3_threshold = 24 needs no
adjustment. The correct finger clears it by ~40x, so the threshold was never
the problem.

Full stack now works end to end — fprintd-verify returns verify-match, and
pam_fprintd authenticates sudo.

One thing this PR does not cover: enrolment on this sensor also hits the 180s
DEFAULT_TEMP_HOT_SECONDS cutoff and aborts partway. I opened #39 for that
separately; it touches different lines and does not conflict with this.

[finikorg]

Follow-up: I've now tested the crop_frame contrast normalisation as well, so
this covers your full goodix511.c change rather than just the algorithm
line from my previous comment.

It measurably works, and it fixes a problem I had been working around by hand.

Without it, enrolment yield tracked how moist the fingertip was. Across one
session it degraded as repeated presses dried the finger:

condition	presses	accepted	yield
moist	44	20	45%
drying	66	12	18%
dry	44	2	5%

Deliberately breathing on the finger before each press took it from 2 accepted
presses in two minutes to 14 — roughly a 7x swing purely from skin moisture. I
had written that up as an unavoidable quirk of the sensor.

With your normalisation, same finger, same sensor, deliberately dry:

21 presses -> 20 accepted, 0 rejected -> enroll-completed
fprintd-verify -> verify-match

95% yield dry, versus 5% dry before. So this is not a marginal improvement:
it removes the moisture dependency altogether, which for day-to-day use matters
about as much as the algorithm fix does.

Makes sense given the sensor output — the raw frames occupy a narrow band of
the 0–255 range, and SIFT keypoint detection is contrast-sensitive. Stretching
per frame rather than assuming a fixed range is the right call, since the band
shifts with skin condition.

Both hunks in this PR are now verified independently on Ubuntu 24.04 / MateBook D
(27c6:5110), separate from the original Debian report. Full stack works:
fprintd-verify matches, pam_fprintd authenticates sudo, and the Cinnamon
lock screen unlocks by fingerprint.

I have not tested the meson.build / doctest changes — they are Debian build
fixes and Ubuntu 24.04 builds fine without them.