Add Voxtral transcription skill

[fazpu]

Summary

• Add a docs-only voxtral-transcribe skill for generated-audio transcription evals with Mistral Voxtral.
• Document stable transcript artifact handoff under eval_artifacts/transcripts/ versus per-check evidence under EVAL_BANANA_OUTPUT_DIR.
• Add reusable eval patterns for same-check validation, stable handoff, dual write, deterministic marker checks, JSON shape checks, and semantic judge checks.
• Update the README skills section to list the repo-local skills, including voxtral-transcribe.

Notes

• This PR intentionally adds no helper script, no mistralai dependency, no live API tests, and no eval-banana runtime changes.
• The skill name uses hyphen-case for skill-creator compatibility, despite the older gemini_media_use underscore precedent.
• The approved implementation plan predated the current main check-type refactor. Current main supports deterministic and harness_judge; llm_judge is rejected as renamed and task_based has been removed. The new skill examples therefore target current repo semantics instead of adding stale examples.
• The old eval-banana distribute-skills dry-run command has also been removed on current main; README now directs users to npx skills.
• No live Mistral transcription calls were made, and no secrets were read or printed.

Mistral Docs Checked

• Voxtral Mini Transcribe 2 model card: https://docs.mistral.ai/models/model-cards/voxtral-mini-transcribe-26-02
• Audio transcription overview: https://docs.mistral.ai/studio-api/audio/speech_to_text
• Offline transcription guide: https://docs.mistral.ai/studio-api/audio/speech_to_text/offline_transcription
• Audio transcriptions API reference: https://docs.mistral.ai/api/endpoint/audio/transcriptions

Relevant current findings:

• voxtral-mini-latest is the documented transcription model alias; the model card lists voxtral-mini-2602 and pricing around USD 0.003/min.
• Current Python examples use client.audio.transcriptions.complete(...); official surfaces still vary between from mistralai import Mistral and from mistralai.client import Mistral.
• API reference lists file, file_url, file_id, language, temperature, timestamp_granularities, diarize, and context_bias.
• timestamp_granularities supports segment/word in the current docs and is documented as incompatible with language.
• Current overview says long-audio support is up to 3 hours; the skill tells consumers to live-verify long podcast behavior before relying on it in CI.

Validation

• python /Users/jpuc/.codex/skills/.system/skill-creator/scripts/quick_validate.py skills/voxtral-transcribe - pass
• uv run ruff check src tests skills - pass
• uv run ruff format --check src tests skills - pass
• uv run pyright src tests skills - pass
• uv run pytest tests -v --tb=short - pass, 115 passed
• make all-check - pass, 115 passed
• git diff --check - pass
• find skills/voxtral-transcribe -type l -print - no symlinks
• Secret scan found only placeholder/runtime env-var references and an existing README GitHub Actions secret expression.

Distribution dry-run caveat:

• Config discovery from this nested checkout resolves to /Users/jpuc/code/moje/paper_digest, with no skills_dir attribute on current config.
• uv run eval-banana distribute-skills --dry-run and uv run eval-banana distribute-skills --target-agents codex --dry-run now fail with No such command 'distribute-skills' on current main, because skill installation moved to npx skills.