[FEATURE] tools/compare_renders.py: compare two render builds, model by model

[squatchlr]

Adds tools/compare_renders.py, for reviewing a speed-up: does the faster build still produce the same audio, model by model?

python3 tools/compare_renders.py build-base/tools/render build-candidate/tools/render

It renders every model (default example_models/*.nam; .wav impulse responses work too) with both binaries and prints, per model:

• EXACT: byte-identical;
• NEAR: error-to-signal ratio below -100 dB and no sample off by more than 1e-4;
• DIFFERENT: anything else, with the first differing sample and the error.

It exits 0 if no model differs, 1 if one does, and 2 if a render fails.

---
title: Reviewing a speed-up
---
flowchart TD
  accTitle: Reviewing a speed-up
  accDescr: Build the base and the candidate, run compare_renders.py on both render binaries, then read the speed numbers only if every model is exact or near.

  n_build["build base and candidate"] --> n_cmp["compare_renders.py base cand"] --> n_q{"every model<br/>EXACT or NEAR?"}
  n_q -- yes --> n_speed["look at the speed numbers"]
  n_q -- no --> n_ask["which model, which sample?"]

Loading

The test signal is generated at each model's own sample rate, with an odd length so the last block is always short. It is a sweep, noise, silence and a full-scale square. Standard library only, so there is nothing to install. --self-test checks the comparison itself: EXACT, NEAR and DIFFERENT thresholds, one wrong sample, a NaN, and truncated renders.

render writes output channel 0 only, so only that channel is compared. Comparing an -O3 build with an -Ofast build of main: two models are EXACT, the rest are NEAR (worst -119 dB, largest sample difference 1.8e-5).