[cekrauseee]

The notes repository contains the approved multilingual article and the complete ElevenLabs v3 narration pipeline. It keeps one Markdown source per locale, preserves recognized audio tags for speech generation, derives a clean spoken projection for alignment, and produces one final audio/alignment pair per locale.

Changes

• Make Eleven v3 the default model with Natural stability (0.5) while preserving legacy model controls.
• Recognize a controlled set of audio tags, include them in generation identity, and keep them out of spoken-text hashes and word alignment.
• Split long notes at paragraph, sentence, and Unicode-safe boundaries within the model limit.
• Persist each completed block response so interrupted runs can resume without repeating completed paid calls.
• Validate and assemble MP3 frames, then shift block timestamps onto the actual assembled timeline.
• Update the publishing contract and regression coverage for tags, Unicode offsets, segmentation, recovery, MP3 assembly, and provider payloads.
• Keep the versioned empty .notes/manifest.json; local generated media remains ignored and is not part of the publication catalog.

Verification

• npm run check passed: formatting, lint, typecheck, 27 tests, and note validation.
• Local narration was generated for the note in English, Portuguese, and Japanese. The resulting MP3 and alignment artifacts were checked; durations are 255.09s, 316.24s, and 396.51s respectively.
• No Blob upload or public manifest publication was performed.
• Audio quality, pronunciation, and transitions still need a listening review.