[BUGFIX] Add activation output to residual when layer1x1 is inactive

[andrenepomuceno]

Fixes #341

Problem

When layer1x1 is inactive, detail::Layer::Process copied the layer input straight to _output_next_layer, so the residual path dropped the layer's activation. The trainer computes residual = x + layer_output, and layer_output is post_activation when there is no layer1x1 (nam/models/wavenet/_layer_array.py, Layer.forward, ~L755–773 in neural-amp-modeler 0.13). Any model exported with "layer1x1": {"active": false} therefore rendered differently from the trainer.

Change

• NAM/wavenet/model.cpp: with layer1x1 inactive, _output_next_layer = input + _z.topRows(bottleneck). This is the activation output after the optional activation_post_film, the same tensor the trainer adds.
  • Eigen path: one expression.
  • NAM_USE_INLINE_GEMM path: the memcpy becomes a per-column add. It uses _z's real row stride because _z has 2 * bottleneck rows under gated/blended activations.
• NAM/wavenet/detail.h: updated the doc comments that described the old identity residual.

No new validation is needed. detail::Layer already throws when layer1x1 is inactive and bottleneck != channels, and test_layer1x1_inactive_bottleneck_mismatch covers that. Neither the A2 fast path (it requires an active layer1x1) nor the slimmable WaveNet (it reuses detail::Layer) needs changes.

Tests

• test_layer1x1_inactive: the expected value now comes from the trainer's formula (1 + ReLU(2) = 3). It used to be 1, which asserted the bug.
• New test_layer1x1_inactive_gated: a gated layer with per-frame values and constant gates. It checks that only the gated (top) rows of _z reach the residual, which exercises the strided copy in the inline GEMM build.
• New test_layer1x1_inactive_full_model: a two-layer, single-channel WaveNet loaded through get_dsp from a JSON config with layer1x1 inactive. The expected output is computed in the test from x1 = x + ReLU(a0*x + b0 + m0*x), so the second layer (and the output) depend on the residual.

Cross-check: a randomly initialised neural-amp-modeler 0.13 WaveNet with inactive layer1x1 and a bounded Hardtanh, rendered with tools/render, matches the PyTorch forward pass (pad_start=True, 4096 samples) to a max abs error of 7.5e-8 with this PR and #342 (Hardtanh bounds) applied, down from 1.9e-3 on main.

run_tests passes in both the default and -DNAM_USE_INLINE_GEMM builds (g++ 11, Linux). Before the fix, the updated test_layer1x1_inactive fails in both builds.

Note

This changes the output of existing .nam files that have layer1x1 inactive: they now match the trainer, but the difference can be audible for models that were only ever heard through the core.