Fix one-bin frequency shift in get_response (fixes #3)

[jankoslavic]

Summary

Fixes #3 — get_response(domain="f") synthesised responses using an FRF shifted by one frequency bin relative to get_FRF_matrix.

Root cause

get_response built its frequency vector as

len_freq = exc.shape[1] // 2 + 1
freq = np.arange(1, len_freq+1, 1) * (sampling_rate / exc.shape[1])  # avoid zero frequency

i.e. [1·df, 2·df, …], starting at df. But EXC = np.fft.rfft(exc) has its bins at [0, df, 2·df, …]. The response loop then multiplies matrix[..., i] (the FRF at (i+1)·df) by EXC[:, i] (the spectrum at i·df), so the synthesised transfer function is shifted up by one bin: H_used[k] == get_FRF_matrix[k+1].

Fix

Use the aligned rfft grid, which includes the 0 Hz (rigid-body / static) term:

freq = np.fft.rfftfreq(exc.shape[1], 1/sampling_rate)

The 0 Hz FRF is kept where it is defined (grounded systems → inv(K)) and left at zero where the impedance matrix is singular (unconstrained / free systems, frf_method="f"), so get_response never fails at 0 Hz while reproducing get_FRF_matrix at every other bin:

frf_matrix = np.zeros((self.n_dof, self.n_dof, freq.shape[0]), dtype="complex128")
frf_matrix[:, :, 1:] = self.get_FRF_matrix(freq[1:], ...)
try:
    frf_matrix[:, :, :1] = self.get_FRF_matrix(freq[:1], ...)   # 0 Hz rigid-body term
except np.linalg.LinAlgError:
    pass                                                        # singular -> left at 0

The "f" and "t" domains share this handling.

Verification

before:  get_FRF_matrix vs get_response FRF, full band  max|Δ| = 2.6e-3   (one-bin shift)
after :  get_FRF_matrix vs get_response FRF, full band  max|Δ| = 0.0      (0 Hz included)

• Grounded systems: get_response(return_matrix=True) equals get_FRF_matrix at every bin, DC included.
• No crashes at 0 Hz for any boundary condition — both/left/right/free, both frf_methods, both domains all run and stay finite (previously free + "f" raised LinAlgError at DC).
• domain="f" and domain="t" responses agree.
• pyLump test suite: 4 passed. test_response now aligns its reference grid and excludes the 0 Hz bin from the comparison (by removing the time-domain mean, which zeros exactly the DC bin).

Found while adding pyLump-based cross-validation tests to pyFRF.

[jankoslavic]

Closing for now — nothing actually depends on this. The pyFRF tests that surfaced it compare against get_response(return_matrix=True) (the exact FRF used to synthesise the data), so they pass with or without this change. The underlying one-bin inconsistency between get_response and get_FRF_matrix is left documented in #3 for whenever it's worth addressing.

[jankoslavic]

This does not make sense.