Skip to content

feat: version conversation extraction and independent text publications - #2

Draft
nichinichisou0609 wants to merge 7 commits into
mainfrom
codex/common-talk-translations
Draft

nichinichisou0609 wants to merge 7 commits into
mainfrom
codex/common-talk-translations

Conversation

@nichinichisou0609

@nichinichisou0609 nichinichisou0609 commented Oct 25, 2026 •

Copy link
Copy Markdown
Contributor

Ordinary conversations between multiple characters were omitted from the dialogue export. This adds their own corpus and voice selection, plus the separate communication motion scripts used while characters gather and wait.

The new source, export and depot command groups introduce versioned, content-verified tasks over the existing pack store, following the source/task/result/product separation used by nnnotes. Existing flat exports remain available. Task results participate in integrity checks and GC retention.

Text catalogues maintain stable IDs and source hashes across updates. Official regional matching checks script, speaker, participant and display structure. Runtime source shards and translation publications are separate: editing a translation leaves resource snapshot inputs unchanged, and consumers can admit unchanged lines across source revisions. Compact search shards support translated titles and dialogue search without per-script downloads. An independent official-text preparation command reuses the existing uploader and switches only the translation pointer; human editing uses the companion NextTrans integration. This PR does not deploy content.

An independent font catalogue publishes exact base-font coverage, supplemental open fonts and their licenses. The same uploader supports its immutable objects and explicit latest-pointer update. Text edits change neither font objects nor resource snapshots; translators only edit text.

Tutorial scripts, voice-bank selection, canonical placement poses and validated introductory pose configuration supply the companion NPC runtime. Photo-shot UI decoders expose its view, presenter and animation bindings; dynamically loaded call effects use separate reachable object closures. Character sidecars add authored head axes, renderer bounds and each actor's exact controller clip identities; the registry joins its original character identity and gender without inferring them from model names.

Validation:

  • Latest CI on the checked-out commit: Python 1,365 passed / 19 skipped on Ubuntu and 1,359 passed / 25 skipped on Windows. Private-source tests remain opt-in.
  • Wheel/sdist build, isolated installation, CLI and pack/history smoke checks passed on both CI platforms after the font catalogue addition. Runtime font rendering remains a companion-stage gate.
  • Caller-supplied source validation: 117 ordinary multi-character conversations, 553 requested voice cues decoded from 117 banks, and 30 communication scripts / 3,705 steps. No source assets or derived corpus are included.
  • Repeated task reuse, relocation, changed-input rejection, corrupt-result retention checks, and translation-only changes preserving source shard bytes passed.
  • Final text product: editing one line preserved all 14,195 source files byte-for-byte. It reused 11,688 translation objects and required five new/updated files totaling 374,581 uncompressed bytes, including the affected search shard and publication roots.
  • Tutorial placement extraction consumed 11 source assets / 56 participant poses with no unread bytes; each photo call effect resolved to one prefab and one particle system with no unsupported objects.
  • Font preparation and upload selection now independently validate the complete referenced closure, canonical revision, strict UTF-8, digest paths, metadata and every actual local upload source. Fifty-two focused font/HTTP checks passed. The supplied supplementary font covers 34,047 additional codepoints without changing the base font or WASM. This is local preflight; immutable remote uploads must succeed before switching the latest pointer.
  • All 36,270 controller references across 31 supplied character models joined the existing motion index by exact source identity, with no missing or duplicate matches. The additive export preserved the existing runtime files.

Reviewed migration fingerprint changes cover the animation CLI's default output directory and additive character sidecars. Character geometry, skin and animation coordinate writers are unchanged.

Draft while the companion runtime, embedded stage and host integration complete end-to-end validation.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant