Add downstream task evaluation workflows

[Tinnci]

Goal

Evaluate whether enhancement actually helps downstream tasks instead of only improving audio metrics.

Scope

• Add a downstream eval interface for comparing low-quality input vs enhanced output.
• Track ASR WER/CER before and after enhancement.
• Track speaker embedding similarity so enhancement does not destroy speaker identity.
• Track VAD or endpoint accuracy where labeled fixtures are available.
• Leave room for keyword spotting accuracy and TTS/voice-conversion preprocessing checks.
• Make downstream models optional and gated.

Acceptance criteria

• Downstream eval output records baseline input score, enhanced score, delta, evaluator version, and dataset identity.
• ASR WER/CER has a documented adapter plan or first optional adapter.
• Speaker similarity and VAD/KWS are represented in the schema even if initially deferred.
• Documentation warns that SR improvements that worsen WER or speaker similarity are regressions for some use cases.
• Default tests use mocks or tiny generated fixtures only.

[Tinnci]

This was generated by AI during triage.

Implemented and verified the #40 downstream task evaluation slice.

Evidence:

• 14053ab adds audio-super-res eval downstream.
• Builtin transcript-error-rate evaluator computes ASR WER/CER for precomputed low-quality baseline transcripts and enhanced-output transcripts against a reference transcript.
• Downstream records include dataset identity, evaluator/version, baseline input score, enhanced score, delta, task, status, and metadata.
• Speaker similarity, VAD/endpoint, and keyword spotting are represented as planned optional/gated adapters with runtime and install guidance.
• Documentation warns that SR improvements that worsen WER/CER are regressions for ASR preprocessing scenarios.
• Default tests use JSON fixtures only and do not install or run ASR/speaker/VAD/KWS models.

Validation:

• Local gates passed: pixi run lint, pixi run format-check, pixi run typecheck, pixi run test, pixi run build, pixi run metadata-check, pixi run wheel-check.
• Colab CLI validation passed on 14053ab: fresh clone, uv pip install --system -e . pytest, tests/test_eval.py, audio-super-res eval downstream, and gated --evaluator speaker-similarity failure path.
• GitHub Actions CI and Security both passed for 14053ab.