Tests: nothing checks that sound is actually recovered; add a synthetic vibrating-surface test with known ground truth #20

Severity: Medium. The current tests check shapes, finiteness and range. The smoke tests feed random-noise frames, where there is no signal to recover, so a change that broke recovery would still pass. That's how #13, #14 and #17 went unnoticed.

Weak or misleading tests

• test_constant_input_silent asserts only shape and finiteness. It doesn't check that the output is silent.
• test_passband_preserved asserts std > 0.01 on output that is normalised to [−1, 1], so it passes for any non-constant signal.
• test_stopband_attenuated compares standard deviations of two outputs, each normalised to its own peak. It passes because of how the 200-sample edges behave, not because the filter works.
• The input-validation tests pass a file dummy.avi that doesn't exist, so any run that exits non-zero passes, even for an unrelated reason.

Proposed ground-truth test

This generator was used in the audit. It makes exact sub-pixel motion via the Fourier shift theorem, so recovery can be scored against the known signal:

def texture(n=128, seed=0):
    t = ndimage.gaussian_filter(np.random.RandomState(seed).rand(n, n), 2.0)
    return 40 + 170 * (t - t.min()) / (t.max() - t.min())

def shift_img(img, dx, dy):
    H, W = img.shape
    ky, kx = np.fft.fftfreq(H)[:, None], np.fft.fftfreq(W)[None, :]
    return np.real(np.fft.ifft2(np.fft.fft2(img) * np.exp(-2j*np.pi*(kx*dx + ky*dy))))

# frames_k = shift_img(tex, a*s[k]*cos(φ), a*s[k]*sin(φ)) + noise,  s = 440/660/250 Hz tones, fps = 2200
# Feed through a FakeCap(frames) exposing read()/release() into extract_audio / extract_audio_gpu

Suggested assertions:

1. Horizontal, vertical and 45° motion at 0.05 px: correlation with the truth ≥ 0.99 (vertical currently gives 0.85-0.91, see Sub-band alignment ignores sign: anti-phase orientations are misaligned or cancel (recovery drops to r≈0.85-0.91 for vertical motion) #13).
2. 0.3 px drift with default settings: correlation ≥ 0.95 (currently 0.26, see With default settings (no filter), slow drift swamps the audio: 0.3 px of drift cuts in-band level by 12 dB #14).
3. 3 px drift with -fl 100: correlation ≥ 0.99 (currently 0.95, see Measuring phase against frame 0 wraps once the surface drifts; use frame-to-frame differences plus a cumulative sum #17).
4. CPU vs GPU-path output correlation ≥ 0.999 (GPU run on CPU torch; see the CI issue).
5. Over- and under-reported frame counts: the output length equals the decodable frames (see GPU path drops the last partial batch when the container over-reports frames; CPU path drops frames when it under-reports #16).

Keep the clips short (1100 frames × 128², about 30 s each on CPU; or 64² for CI speed) and mark them @pytest.mark.slow if needed.