Fix OBR resampler downsampling errors and rate-dependent binaural loudness

[trsonic]

Three DSP fixes in the vendored OBR subtree (src/renderer/obr/obr_capi/obr/obr/ambisonic_binaural_decoder/), each with regression tests. These surface whenever the output sample rate differs from the 48 kHz-native HRIR assets, or when audio is processed in blocks shorter than the resampler's filter state.

1. Polyphase partition bug in the resampler (downsampling gain/anti-aliasing errors)

GenerateInterpolatingFilter derived coeffs_per_phase_ from max(up_rate_, down_rate_), but the polyphase decomposition built by ArrangeFilterAsPolyphase has up_rate_ branches. Whenever down_rate_ > up_rate_ (downsampling), the tail of the interpolation filter was silently dropped; for integer ratios the kept segment ends before the sinc main lobe. Measured gain errors: +0.65 dB (88.2k→48k), −2.47 dB (96k→48k), −35.1 dB (192k→48k), −41.0 dB (176.4k→48k), with correspondingly degraded anti-aliasing (THD+N as poor as −5 dB at 176.4k→48k). The common 44.1k↔48k pair only loses near-zero window-tail taps, which is why this went unnoticed. Upsampling was unaffected.

Fix: partition the prototype filter across up_rate_ phases. Note this increases coeffs_per_phase_ (per-sample cost and state size) for downsampling — the correct cost the truncated filter was not paying.

2. State-buffer addressing for input blocks shorter than the filter state

Process() carries the last coeffs_per_phase_ − 1 input samples between calls in state_. The branch handling short input buffers addressed the buffer relative to state_channel.end(). That was correct in the original Resonance Audio code, where state_ was allocated at exactly coeffs_per_phase_ − 1 frames, but this port preallocates state_ at kMaxSupportedNumFrames, so end()-relative writes landed in a dead region while the convolution reads the state from the front. Any block-based use with blocks shorter than the filter state produced corrupted output (near-silence with glitches at block boundaries).

Fix: address the live region relative to begin(), matching the read path and the long-input branch.

3. HRIR filter gain not compensated for sample-rate conversion

CreateShHrirsFromWav resamples the SH HRIRs to the target rate with a waveform-preserving resampler and uses the result directly as convolution filters. Those two operations have different invariants: the gain of the filter realized by discrete convolution is proportional to the rate at which the underlying continuous impulse response is sampled, so the uncompensated result scales the rendered binaural level by target_rate / native_rate at every frequency. With the 48 kHz-native bundled assets, rendering at 96 kHz came out +6.02 dB loud and 44.1 kHz −0.73 dB quiet (measured as exactly 2× and exactly 44100/48000 on the DTFT of the loaded filters).

Fix: scale the resampled HRIRs by wav_rate / target_rate so rendered loudness is invariant to the output rate.

Tests

• resampler_test: unity passband gain (±0.1 dB) for a 1 kHz tone across seven rate pairs (fails on the old code at e.g. amplitude 0.018 for 192k→48k); anti-aliasing assertion for 96k→48k and 88.2k→44.1k (the tone previously leaked at −11 dB); ChunkedProcessingMatchesOneShot verifying block-based processing is sample-exact against one-shot for both directions, including blocks shorter than the filter state.
• New sh_hrir_creator_test: cross-channel response-norm invariance over 44.1/88.2/96 kHz at five frequencies, per-channel invariance across asset variants (orders, profiles, both ears), native-rate load determinism, and resampled length vs. rate ratio.
• All pre-existing OBR CMake tests pass.

[jingbo-marquis]

looks great.

[felicialim]

is it worth adding a test where the downsampling factor is not an integer?

[trsonic]

Good idea. Added in the 50846a162635a9688794a1983be607a56cb5df9c commit.