[RamazTs]

🎯 What problem does this PR solve?

• ACE-Step normalizeLoudness: false did not return raw output. The option reaches the engine, but the addon then rescaled every non-edit result to a 0.9 peak before the int16 conversion, whatever the option said. The flag only switched off the engine's percentile pass.
• The same rescale was applied to lego stems. The engine never normalizes a stem, so that it keeps its mix gain relative to the source; the addon raised or lowered each stem to a 0.9 peak.

📝 How does it solve it?

• The gain decision moves out of AcestepModel::generate into addon/src/model-interface/acestep/OutputGain.hpp. It follows the engine's rule: only a generation that asked for normalization is rescaled. normalizeLoudness: false, edit plans and lego stems keep gain 1.0.
• A normalized generation is unchanged: still scaled to a 0.9 peak, still left alone when near-silent.
• With normalization off, samples the VAE leaves outside [-1, 1] now saturate at int16 full scale. That is what raw output means at an int16 boundary; the README and the option's doc comment say so.
• The cpp-tests lane now runs npm run test:cpp (the addon's own gtests, this one and the budget and runtime-option tests) after the native build's caches are saved and before the fuzz build; it only printed a stub before. The workflow policy test expects the new step.
• No JS API change. index.d.ts is regenerated for the doc comment. CHANGELOG entry under Unreleased, no version bump (package is at 0.10.0, released today).

Scope note: QVAC-27588 also lists ACE-Step backend / lmBackend and MiniMax coremlConditionerPath. Those are already implemented in #5495, so this PR leaves them alone. It touches a different part of AcestepModel.cpp and should merge cleanly on either side of it.

🧪 How was it tested?

• New gtest addon/tests/test_output_gain.cpp (6 cases: normalized generation, normalization off, lego stem, edit, near-silent and empty input, quiet input), added to audiogen_ggml_tests. Compiled and run locally together with the existing budget tests against an existing vcpkg tree: 11/11 pass.
• AcestepModel.cpp syntax-checked with the package's compile flags against the engine header at the pinned speech-cpp commit; clang-format clean.
• npm run lint:js, npm run lint:ts, npm run test:types (typecheck, d.ts, generated-wrapper check) pass; JS unit tests 185/185; test:node 47/47.
• Not done locally: a full native build and a real generation. The machine did not have the disk for a clean vcpkg build, so the changed code path has not produced audio here; CI's native build and integration lanes are the first real run. test:package was not completed locally (it downloads model files).

[github-actions]

Review Status

Current Status: ✅ APPROVED
Approvals so far: Team Lead: 1, Member: 1

[github-actions]

License compliance — clean

No new dependency license findings in this PR.

Warn-only (shadow) mode — this check does not block merges yet.

Updated automatically by the canonical license compliance workflow.

NOTICE presence (advisory)

Missing NOTICE (advisory, does not block):

• ./docs/website
• ./packages/asr-ggml/benchmarks/server
• ./packages/fabric/test/integration
• ./packages/llm-llamacpp/benchmarks/performance
• ./packages/llm-llamacpp/benchmarks/server
• ./packages/inference-addon-cpp/mobile
• ./packages/sdk-e2e
• ./packages/embed-llamacpp/benchmarks/performance
• ./packages/embed-llamacpp/benchmarks/server
• ./packages/vla-ggml/sim/server
• ./.github/actions/release-merge-guard
• ./.github/bare-toolchain

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no new actionable issues were found.

Summary

ACE-Step now keeps the engine's gain when normalizeLoudness is off or the output is a lego stem.

• ACE-Step keeps raw gain for edits, stems, and unnormalized audio.
• The AudioGen C++ workflow runs addon tests before its fuzz checks.

Review checks were based on source inspection; no native build or audio generation was run.

_{Reviews (4) · Last reviewed commit: "Merge branch 'main' into fix/QVAC-27588-..." · Reviewed by Greptile}

[github-actions]

Workflow security (shadow mode)

zizmor found 782 finding(s) in .github/ (highest severity: high). This check is warn-only and does not block the merge.

Findings are annotated inline on the changed files and listed in the job summary.

Reproduce locally:

pipx run zizmor==1.27.0 --offline .github/