[haaziq-sl]

CPU and speech optimization currently require custom orchestration and scorer scripts, while agents repeatedly reload logs and saved-run context. This adds reusable CPU/media benchmarking and speech-quality tools, plus compact evidence inspection and validated run resumption.

Changes

• Require fresh web research for baselines and tuning recipes on the exact model and hardware. Engine lists are examples with no fixed count; recheck past surveys and platform exclusions, then tune and measure locally with versions and workloads pinned for each comparison.
• Support fresh-process audio/image rounds, stable audio identities across port changes, TTS intake fields, CPU diagnostics and thread/affinity screens. Reports retain environment and memory context and reject incomplete round coverage.
• Add opt-in evidence summaries, post-timing request/token accounting and read-only run resumption with policy/evidence validation, including compatibility with older media intakes. Full evidence remains accessible.
• Add local Faster Whisper, UTMOSv2 and SpeechBrain ECAPA adapters, with WER/CER, naturalness and optional voice-reference checks. Scorers run in a separate CPU environment, bind observations to registered policies and saved PCM, and retain exact-output and separate listening requirements.

Addresses #5, #6, #7, #8, #9 and #10.

Validation

• Runtime/scorer changes: full suite 1,971 passed, 90 skipped, including installed-wheel and MCP checks.
• Baseline-selection instruction update: 386 passed, 1 skipped across skill, workflow, plain-language and OpenCode checks.
• Real CPU execution of all three scorers and the complete policy → saved-audio replay → scoring → quality → report workflow, with network calls blocked.
• Missing/modified PCM and mismatched score identities were rejected; identical inputs produced matching observations.
• The real-model smoke used one public synthetic sample. Its quality verdict correctly remained UNPROVEN; this was not a full KittenTTS optimization or listening study.

Release follow-up

The package version and plugin install pins remain at 0.1.0. Users need a new PyPI runtime release and coordinated launcher/install-pin updates to receive the new operations. This PR does not publish a package.