Studio 2.2.5: ILC/ILA benchmarks, interactive equations and simpler installs

[gaochangw]

Studio now supports explicit PA-input creation/removal, interactive LaTeX equations and ILC-assisted ILA DPD experiments. Training-only waveform learning fits the transferable predistorter; a separately labelled Ideal reference uses test-waveform feedback. Results keep those signals, scores and provenance separate.

Custom OFDMA defaults to shared channel settings, next-step actions move above configuration, metric calculation methods get descriptive labels and LaTeX definitions, and About stops loading GitHub activity. Default installation includes PyTorch, pywebview and platform bindings; the uv guide starts from environment creation and covers automatic backend selection and SSH forwarding. The reported connection refusal was confirmed to be Mac-to-Linux loopback misuse; browser fallback now explains that case and is regression-tested as a live service.

Validation completed locally: 231 frontend tests; strict types/lint/build/contracts; 24 Chromium journeys/accessibility checks; eight actual CPU PA/DPD workflow jobs; default ILC runs on CPU and CUDA; all nine interactive formula catalogs; tenant-scoped input removal/Undo; clean default installation; packaging/golden/ILC integration tests; strict documentation build and refreshed GUI screenshots. The final full Python regression and Linux/macOS/Windows CI are required before merge.

ILC remains experimental and surrogate-based. Its Ideal reference is waveform-specific; configured target NMSE is not guaranteed. Polynomial arithmetic uses CPU complex128 while PA replay uses the selected accelerator. Native-window and MPS hardware coverage remain explicitly limited in the support matrix. Release notes and validation records are under docs/releases and docs/performance/studio-2.2.5.

[gaochangw]

Reviewed commit 9e20cd5. No blocking findings.

Checked training/test separation in ILC/ILA, fixed target/peak handling, monotone backtracking, cancellation, repeat evaluation and waveform provenance. The fitted DPD and feedback-optimized Ideal test waveform retain distinct metrics and signal positions. CPU complex128 polynomial buffers remain on CPU when the PA cascade moves to an accelerator.

Checked the KaTeX trust boundary and adversarial coverage, parameter highlighting after rerenders, per-channel validation/legacy configuration compatibility, tenant-scoped remove/restore with intact paired artifacts, and public ILC compute caps. Default dependencies use platform markers; hardware-aware wheel selection is documented as a uv feature, with drivers and native-window coverage distinguished.

Validation: full local Python regression 909 passed / 19 extended deselected; frontend 231 passed; Chromium journeys/accessibility 24 passed; real generator/PA/DPD CPU workflows; default ILC on CPU, CUDA and the production GPU image; packaged GUI and isolated default installation; strict docs/contracts/lint/build. Linux, macOS and Windows default-install CI has passed. Final merge remains contingent on the remaining PR CI jobs.