Make renderer gain slew callback-partition invariant

[dissonance-git]

What changed

• keep ChannelState gain slew at one constant sample-time rate across callback boundaries
• clamp only after the target is actually reached instead of dilating the last gain fraction to the callback edge
• carry the same start/rate/end segment into binaural rendering, so the headphone path no longer applies block-end gain as a callback-rate step
• add an adversarial regression that renders one continuous fade under two radically different callback partitions and requires sample-equivalent output

Why

Host callback boundaries are transport artifacts. The current slew computes the final partial step as remaining_delta / sample_length, so a callback that straddles the target changes the audible gain trajectory. In binaural mode, the block-end gain is also applied to the whole block, making the trajectory even more directly callback-size dependent.

The regression uses the direct LFE headphone route to avoid HRTF/reverb differences and compares the same 1200-sample fade under 40-sample blocks versus an irregular partition that crosses the 960-sample slew endpoint inside a 273-sample callback.

Validation

The branch is based directly on upstream b1b78a89bb83417d7ff6ae73a617983269bd12a8. This PR is opened as draft so upstream CI can run the repository's format/build/test gates before it is marked ready.

[mgth]

Thanks — the fix is sound and the adversarial partition test is exactly the kind of regression we want pinned. I ran the branch locally on top of b1b78a8: the new test passes, but two golden null tests fail, which is the expected consequence of the intended trajectory change:

• null_binaural_kemar — peak residual −24.8 dBFS, RMS −44.6 dBFS (the binaural path moving from block-end constant gain to per-sample interpolation)
• null_crossover_bands — the speaker path's end-of-ramp clamp behaviour

Two requests before marking this ready:

1. Re-bless the goldens with OMNIPHONY_BLESS_GOLDENS=1 cargo test -p renderer and quote the printed residuals in the PR description, as the harness instructs. CI runs the full suite, so it will stay red until then.

2. Make GainSlew::at() branchless. It sits in the per-sample hot loops of the speaker stage, and this repo's convention is to keep per-sample work branch-free. The sign test on step is loop-invariant, so cheap in practice, but it can block auto-vectorisation. Since the ramp is monotonic from start toward end, hoisting the bounds once per block does it:

   let lo = self.start.min(self.end);
   let hi = self.start.max(self.end);
   // per sample:
   (self.start + self.step * sample_idx as f32).clamp(lo, hi)

   which compiles to fma + minss/maxss. Same pattern applies to the two inlined gain computations in binaural/mod.rs (hoist lo/hi out of the for s in 0..span loops).

One nit, no change needed: with ramp_samples <= 0.0 the new code jumps to the target at the first sample instead of ramping over one block. Unreachable today (GAIN_SLEW_SECS is a positive constant), just noting it's a deliberate semantic.