Add Audacity eval harness and regression reporting

[sehawq]

Summary

• Add a task-based evaluation harness for Audacity CLI with discovery, reporting, and baseline regression checks.
• Introduce 4 stdlib-only eval tasks:
  • Project roundtrip (create → save → open → info)
  • Track + clip flow (add track, add clip, split)
  • Effects registry validation (normalize/fade_in)
  • WAV export (render_mix to file)
• Add eval command to the CLI with --out, --baseline, --update-baseline, and --fail-on-regression.
• Document eval usage and outputs in Audacity README.
• Add pytest coverage for task discovery, report generation, and baseline regression detection.

Why

• Provides an objective, repeatable quality signal for release readiness.
• Detects regressions between versions using a baseline comparison.
• Keeps MVP self-contained and portable (stdlib only, no external dependencies).

Implementation Details

• New eval runner under audacity/agent-harness/cli_anything/audacity/eval/.
• Task modules live in audacity/agent-harness/cli_anything/audacity/eval/tasks/.
• Reports:
  • eval_report.json (machine-readable)
  • eval_report.md (human-readable)
  • artifacts/ for generated outputs
• Baseline regression rules:
  • Pass → Fail on any task is regression
  • Overall success_rate decrease is regression

Known Limitations

• Eval tasks are Audacity-only (other harnesses not yet wired).
• Baseline comparison is MVP-level (status + success_rate only, no metric thresholds).
• No CI integration in this PR (manual execution).
• SoX-backed tests require a local SoX binary on PATH for full E2E coverage.

Future Work

• Add metric thresholds per task (e.g., duration, RMS/peak bounds).
• Add CI job to run eval and gate merges/releases.
• Expand eval harness to other CLIs (Blender, GIMP, etc.).
• Add YAML/JSON task definitions for non-code task authoring.

Tests

• python -m pytest -v