fix: clamp speaker mix weights to [0,1] to keep the frame mix convex

[KakaruHayate]

Problem

DiffSingerSpeakerMix.Create accumulates each mix:<suffix> sample with only a NaN check — the value is never clamped to the track's declared range:

double v = f < sampled.Length ? sampled[f] : double.NaN;
if (!double.IsNaN(v)) w[f] += (float)v;   // no clamp

The track is declared Continuous(color, 0, 0, 1) and the host UI clamps on draw, but the data layer has no hard contract — the codebase already says exactly this in CombineVariance:

用户值钳到编辑量程：宿主 UI 落笔时钳制，但数据层无硬契约（锚点+默认值合成、跨引擎同 key 复用、手改工程皆可越界）

Negative values are the damaging case. They pull Σ back down to ≤ 1, which defeats the sum > 1 guard below: the weights are not normalized, and the default suffix is topped up by 1 - Σ > 1. The result leaves the convex hull — extrapolation instead of interpolation, i.e. an embedding outside the training distribution.

Reproduced against this repo's master (bcd2396), default speaker Alice, Alice=[1,0], Bob=[0,1], Carol=[.5,.5]:

input	current	in convex hull?
Bob=+0.5	[0.500, 0.500]	✅ correct interpolation
Bob=-0.5	[1.500, -0.500]	❌
Bob=-2.0	[3.000, -2.000]	❌
Bob=2, Carol=-1	[-0.500, 1.500]	❌

The last row is the subtle one: Σ is exactly 1, so the normalization branch never fires.

Fix

Clamp the sampled value to [0,1] at accumulation. The sibling phoneme_mix track — declared with the identical Continuous(color, 0, 0, 1) shape — is already clamped at its sampling site in DiffSingerSynthesisSession:

br[f] = (float)Math.Clamp(double.IsNaN(sampled[f]) ? 0 : sampled[f], 0, 1);

so this only brings the two same-shaped tracks into line.

In-range behaviour is bit-for-bit unchanged: values within [0,1] are unaffected by the clamp, and the normalization / default top-up logic is untouched.

Tests

• Added DiffSingerSpeakerMix.cs to the test project's Compile items (zero external dependencies, same treatment as DiffSingerFrames), so it still runs without the TuneLab SDK checked out.
• New DiffSingerSpeakerMixTests covering existing behaviour (default top-up, over-unity normalization, NaN-as-unedited, short track) plus 6 out-of-range regressions (negative ×3, above-one, mixed-sign, per-frame clamping). The invariants asserted are the ones that actually matter: per-frame weights all in [0,1] and summing to 1, and the resulting embedding staying within the per-dimension convex hull of the speaker embeddings.

I stashed the fix and ran the suite first to confirm those 6 genuinely fail without it (the other 38 pass either way), then restored it for 44/44 — otherwise "weights sum to 1" is the kind of assertion that can silently hold no matter what.

Local run of this repo's CI sequence (plugin / tests / smoke tool, Release): 0 warnings, 0 errors.

---

Found while reviewing an unrelated feature branch in my fork; confirmed independent of it (git diff master...<that branch> -- DiffSingerSpeakerMix.cs is empty), so it's sent on its own.