Davy7/25/2026, 3:54:46 PM

Two deterministic additions so per-target provider behavior is proven
and measured by CI instead of manual experiments:

- A gated test runs a full separation under the PLATFORM-DEFAULT
  execution-provider preference on every cross-target leg. On GPU-less
  Windows runners this exercises exactly the DirectML-attempt ->
  XNNPACK/CPU fallback every user without a dGPU hits (research
  confirmed the DML session requirements are auto-corrected by ORT
  v1.27.1 at EP registration and our architecture matches the
  documented constraints; the fallback fires at session creation,
  never mid-song). On GPU machines the same test exercises real DML.
- The bench harness (spectral-bench/v2) measures the CPU EP and the
  platform default side by side per target, so provider selection
  (#170) is data-driven from every workflow dispatch. Local sample
  (aarch64 macOS): XNNPACK warm 2.67 s/window vs CPU 5.08 s
  (RTF two-stem 0.64 vs 1.31).

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Parentb416cdf