Numerical setImage / operator() tests for each feature evaluator

[vladiant]

Summary

The existing feature tests verified construction, init and feature
counts, but never evaluated a feature on a real image. This MR adds
8 numerical tests that drive the full pipeline of every concrete
evaluator (CvHaarEvaluator, CvLBPEvaluator, CvHOGEvaluator):

init() → setImage() → operator()(featureIdx, sampleIdx)

The tests use synthetic images (constant patches and step edges) and
assert closed-form invariants that hold regardless of which feature
index corresponds to which geometric layout, so they remain meaningful
without hard-coding implementation-specific feature ordering.

Library line coverage rises from 42.4% → 44.4%, driven primarily by
HOGfeatures.cpp going from 28% → 77% (the entire integral
histogram + feature evaluation path was previously untested).

Changes

traincascade/test/test_features.cpp

8 new test cases appended, all following the existing Arrange / Act /
Assert pattern with comments per step:

Evaluator	Invariant exploited	Tests added
Haar	Flat image → variance = 0 → normfactor = 0 → operator() short-circuits to 0.0f for every feature	constant returns 0 (BASIC), vertical edge produces ≥1 nonzero, ALL mode flat → 0 (also exercises the tilted-integral branch in setImage)
LBP	Flat image → all 8 block sums equal cval → all 8 >= cval comparisons true → result 0xFF = 255	constant returns 255, horizontal edge produces ≥1 value <255, two-sample isolation by index
HOG	Flat image → all gradients 0 → bin sums fail > 0.001f guard → 0.0f for every variable	constant returns 0 across all numFeatures × N_BINS × N_CELLS vars, vertical edge produces ≥1 nonzero

Inner loops use a fail-fast accumulator into a single CHECK so the
doctest assertion counter stays sane (291 total, not 170k+).

Why these invariants?

Each evaluator has a property that holds for any uniform input:

• Haar uses a normalization factor derived from the image variance.
  On a constant patch the variance is zero, the implementation
  short-circuits, and every feature returns exactly 0. Any non-zero
  output on a textured image therefore proves setImage actually
  populated the integral image and operator() is using it.
• LBP compares each surrounding 3×3 block sum to the center block.
  Equal sums make every comparison true, producing the maximum bit
  pattern 0xFF. A non-0xFF output on a textured image proves the
  pipeline is working end-to-end.
• HOG computes per-bin gradient histograms; on a flat image every
  bin is zero and the implementation's res > 0.001f guard returns
  0.0f everywhere. A non-zero output on an edge image proves
  integralHistogram, normalization and Feature::calc are wired up.

Using closed-form invariants instead of golden values keeps the tests
robust against benign reordering of generated features.

Test results

[doctest] test cases:  88 |  88 passed | 0 failed | 0 skipped
[doctest] assertions: 291 | 291 passed | 0 failed |
[doctest] Status: SUCCESS!

ctest: 100% tests passed, 0 tests failed out of 1

Coverage impact (traincascade/lib/)

Metric	Before	After
Lines	42.4% (1984 / 4676)	44.4% (2076 / 4676)
Functions	77.5% (200 / 258)	79.1% (204 / 258)
Branches	27.7% (1388 / 5006)	28.8% (1442 / 5006)

Per-file gains:

File	Before → After
HOGfeatures.cpp	28% → 77% ⭐
HOGfeatures.h (inline operator() / Feature::calc)	7% → 100%
haarfeatures.cpp	73% → 75%
lbpfeatures.h (inline Feature::calc)	already 100%, now exercised on real data

The HOG jump is the key win: previously no test ever called
CvHOGEvaluator::setImage, so the entire integralHistogram,
gradient-bin accumulation and normalization paths were dark. They are
now covered on both the zero-gradient short-circuit path and the normal
evaluation path.

Risks / Notes

• All tests use in-memory cv::Mat images — no filesystem I/O, no test
  fixtures.
• Runtime contribution is well under one second.
• No changes to library code.

Checklist

• All 88 tests pass locally (ctest --output-on-failure)
• AAA pattern with comments in every test
• Happy paths and degenerate inputs covered (constant images, edges)
• No new resource files committed
• No changes to library code