Preserve LavaSR digital silence

[Tinnci]

Summary

• preserve strictly zero LavaSR inputs after weight, runtime, and device validation but before model inference
• keep near-silence and every non-zero input on the normal model path
• make eval matrix return a non-zero exit code when any run fails
• make the Colab evidence script reject failed matrices
• record the v0.6.0 T4 evidence and update roadmap/release status

Root cause

The v0.6.0 Colab T4 failure-case matrix passed four of five fixtures. Exact digital silence produced deterministic model output with RMS 0.0001772631 and peak 0.000664089, causing the harness to correctly report silence_hallucination. Five repeated CUDA runs had the same SHA256, while the reference and degraded inputs remained exactly zero. The model output, not degradation, WAV writing, or CUDA variance, was the source.

Fix validation

On the same Colab T4 session, weights, and fixtures, the matrix changed from four passed plus one stability failure to five passed with zero failures. Silence duration drift changed from 0.0026667 seconds to zero.

Checks

• local targeted regression tests passed
• pixi run lint
• pixi run format-check
• pixi run typecheck
• pixi run test: 159 passed, 6 skipped
• remote Colab regression tests: 5 passed
• remote LavaSR failure-case matrix: 5 passed
• git diff --check

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

This PR improves LavaSR failure/silence handling and makes CLI evaluation workflows correctly fail when an eval-matrix run fails.

Changes:

• Short-circuit lavasr-compat inference for exact digital silence to prevent deterministic low-level “texture” output.
• Make audio-super-res eval matrix return a non-zero exit code when any run fails, and enforce the same in the Colab evidence workflow.
• Add/refresh documentation and changelog entries to reflect the updated behavior and validation evidence.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/test_lavasr_compat.py	Adds regression test ensuring exact digital silence stays silent and model load is avoided.
tests/test_cli.py	Adds test asserting eval matrix returns non-zero exit code when failures occur.
src/audio_super_resolution/cli.py	Returns exit code 1 when the eval matrix manifest reports failures.
src/audio_super_resolution/backends/lavasr_compat.py	Adds silence short-circuit in enhance() to preserve exact silence.
examples/colab_lavasr_validation.py	Treats matrix runs as failed if the manifest indicates failures.
docs/ACCELERATORS.md	Records validation/evidence notes for the published tag and silence fix.
ROADMAP.md	Updates release milestone status to completed.
CHANGELOG.md	Documents the fixes under an Unreleased section.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.