feat(LTX25-AUDIO-RESAMPLE): resample the take instead of refusing it, and the window is hann

[localai-org-maint-bot]

feat(LTX25-AUDIO-RESAMPLE): resample the take instead of refusing it, and the window is hann

Closes #2583. Order 7 of .agents/specs/ltx25-completion-scope.md §8, gap A19, which that plan places before A3 and A18 because it gates both.

The gap

Ltx2WaveformToLogMel refused every sample rate but the audio VAE's own (ltx2_audio_vae.cpp:1051-1060), and Ltx2DecodeAudioWav refused it one hop earlier so the message could name the file (ltx2_audio_input.cpp:96-105). Between them, an audio-to-video render only accepted a take the caller had already resampled with some other tool.

Upstream refuses nothing here. waveform_to_mel calls resample_audio before the mel transform (ops.py:44-49) and resample_audio filters whenever the rates differ (ops.py:36-42).

The reason written beside that refusal was wrong, in nine places

The message and eight sibling comments called torchaudio.functional.resample "an arbitrary-ratio polyphase kaiser resampler". It is not one. resample defaults resampling_method="sinc_interp_hann" (torchaudio 2.11 functional/functional.py:1441) and only the kaiser branch (:1386-1391) builds i0-weighted taps; ops.py:40 passes neither that argument nor beta. The window is the same hann this tree already builds for the vocoder's BWE stage — Ltx2HannSincResampleFilter1d is this kernel at orig_freq == 1. What was missing was the rational ratio, not the filter.

All nine are repaired, including the two in ltx25-a2v-audio-input.md, the row that wrote them. polyphase kaiser and kaiser resampler now return zero hits across src, include, tests, docs and .agents.

The port

Ltx2ResampleWaveform mirrors resample at every default: gcd-reduced ratio, lowpass_filter_width = 6, rolloff = 0.99, one kernel row per output phase, convolution with stride orig, and truncation to ceil(new * length / orig). It is called from Ltx2WaveformToLogMel unconditionally, because upstream's rate test lives inside resample_audio (ops.py:38-39); guarding at the call site would have moved that early return off the production path.

Ltx2DecodeAudioWav loses its want_sample_rate parameter. Upstream's decoder has no target rate, and the parameter existed only to raise.

The dtype is f32, and that is the mirror

Upstream's waveform is float32 from the decoder on (decode.py:173-176), encode_audio moves it without casting (audio_vae.py:271), and the kernel is built with dtype=waveform.dtype, so the filter itself is f32 (:1376-1397). Measured against the pinned oracle at four rate pairs:

Filter built in	max abs diff vs the oracle
double, narrowed to float	1.03e-05
float, mirroring upstream	1.79e-07

The double filter is more accurate than upstream and therefore fifty-seven times further from it. It would have passed a gate written around it and read as care. This is AGENTS.md's "a token gate cannot detect a dtype that is too wide" in its exact shape.

The gate is 2.5e-07, twice the measured floor: at 0.0 three of the four arms red at 5.96e-08, 5.96e-08 and 1.19e-07, and the equal-rate arm passes because it is a copy. A double-built kernel fails it by forty.

Tests

Upstream ships none at this pin, so sections 8d and 8e of gen-ltx2-vae-goldens.py EXECUTE upstream's own AudioProcessor — at four ratios chosen for arms rather than coverage (o = 1, n = 1, o = 441 with width = 17, and the equal-rate early return), and through waveform_to_mel at a non-target rate. They ride the generator's existing revision anchor, and regeneration is byte-identical.

Red before, green after. Both reds were the refusal itself:

	Before	After
test_ltx2_vae	THREW at ltx2_audio_vae.cpp:1053	48/48 cases, 3217/3217 assertions
test_ltx2_video	THREW 'high.wav' is sampled at 44100 Hz ...	110/110 cases, 4870/4870 assertions

Reachability. The video case enters through LoadVideoEngine + Generate and pairs a 44.1 kHz take with the same PCM bytes relabelled 24 kHz — the wrong answer the refusal existed to prevent. Deleting the production call site makes those two carry the same latent digest (18285287296143238670) while the unit case over Ltx2ResampleWaveform stays green, which is the whole point of the mutation. The tree was restored byte-for-byte and both suites re-run green.

One assertion was written, run, and removed rather than loosened. audio_latent_absmax reads 1.07194 / 1.07293 / 1.07208 across three genuinely different takes, a 0.1% spread, because the trace's absmax is dominated by the encoder's per-channel statistics. It cannot order them. The reason is recorded where the assertion was, and the values are gated at the VAE level instead.

Also

This file's own scope note claimed the analysis half was not ported. The mel front-end has lived here since the A2V input row, so that sentence was already false; repaired in flow.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5 [claude-code]