v0.1.2 — Stronger evidence, safer reruns, a working first example

This maintenance update makes the existing agent-operated workflow easier to
trust and easier to pick up after a break.

What improved

• Approvals keep their history. Passed stages cannot be silently replaced.
  Explicit replacements retain the old ledger and must keep later approvals valid.
• Textures stay tied to the right mesh. The ledger checks the approved
  topology, exported FBX, baked maps and actual passing texture report by hash.
  Missing records and files with the right name but wrong contents are rejected.
• Failed runs cannot borrow an old success. Blender errors and missing outputs
  stop the run. Character compiles and production bakes keep separate attempts,
  and published asset directories are protected from replacement.
• Windows launchers behave consistently. PowerShell 5.1 receipts, ordinary
  stderr warnings, tool discovery and interpreter selection are handled correctly.
  The AI routes share GPU/ComfyUI preflight checks and retain interrupted attempts.
• Resuming takes less reading. docs/STATUS.md gives the current state;
  docs/HANDOFF.md keeps the history. The docs now include a stage/receipt table,
  an index and clearly separated experimental scripts.
• A fresh clone has something to compile. examples/crate/ contains a tiny
  CC0 mesh, texture and recipe, exercised through a real Blender bake, publication
  and FBX reimport. It is a mechanics example, not AI-generated artwork.

Also fixed: the default mascot rig adapter, custom-workspace geometry preflight,
delegated retopology review, concurrent ledger writes and installed-package checks.
Python dependency updates and private security reporting are now configured.

Updating

Update your checkout, reinstall with python -m pip install -e ".[dev]", and run
scripts\verify.ps1. For a first compile, follow examples/crate/README.md.
Ledger-only users can install the attached wheel and run rac --help.

Existing output directories are retained. Choose a new asset id for another
published derivative, or a new --production-name prod-v3 for another production
bake. Historical ledgers missing the newly required bindings fail audit until
the evidence is supplied and reviewed; approvals are not migrated automatically.

Validated on Python 3.11 and 3.12 with CPU-only tests, isolated wheel installations,
Windows PowerShell runtime checks and a Blender 5.2.1 fixture. No AI inference or
existing asset approval was needed for this update.

v0.1.1 — More reliable checks, safer exports, easier setup

This update makes the existing Codex-driven workflow more dependable, especially
when something fails or you're getting started on another machine.

Fixes

• Missing work can no longer look finished. Asset audits now require every
  expected stage before reporting production readiness. Damaged or incomplete
  records are flagged clearly, and existing evidence is preserved.
• Failed prop exports stay failed. An old report can no longer make a failed
  Blender run look successful. Each compile keeps its own files and logs, checks
  that the new FBX and textures match their receipts, and protects existing outputs.
• The installed CLI works outside the repository. The Python package now
  includes the configuration needed to plan, create and audit asset workspaces.
• Setup checks match what you're trying to run. Check just the ledger,
  geometry, texturing or Unreal route. Missing optional tools won't block an
  unrelated workflow, and JSON output reports failures correctly.
• Local checks match CI. Verification now tests the installed package as well
  as the source, and includes the same tests, lint and syntax checks on both paths.

Getting started

Update your checkout and reinstall the development dependencies:

.\.venv\Scripts\python.exe -m pip install -e ".[dev]"
.\scripts\workflow_doctor.ps1 -Profile ledger
.\scripts\verify.ps1

For ledger-only use, install the attached wheel with python -m pip install <wheel-file> and run rac --help. The full Blender/Unreal pipeline still runs
from the source checkout. Model weights, engines and generated artwork are separate.

Existing work stays put. Use a new asset id when compiling a derivative into
an output directory that already exists. Historical reports remain readable;
this release does not migrate assets or change their approval status.

Validated on Python 3.11 and 3.12, including fresh wheel installs, plus a real
CPU-only Blender export/reimport fixture. No AI inference was needed for this
maintenance release. Visual review and cooked-runtime gates still apply.

The Workshop Update — Textures, Rigging & Real UE Showcase

The Workshop Update

This is a big step forward for Reference Asset Compiler: a textured, animated character inside an atmospheric Unreal Engine workshop, with real engine captures showing the result.

A significant amount of work went into getting the character's face, neck transition and locomotion to this point. This update shares the result—and the checks behind it—without pretending that a single model call did all the work.

What changed

• The workshop is now the main showcase. The README leads with the actual UE 5.8 scene: warm work lamps, a moon and starry sky, a 3D desert vista, and individually placed furniture and props. Three in-scene shots replace the black-background presentation.
• A clearer account of the texture workflow. The demonstrated character uses a separately acquired, image-conditioned AI head fitted to the body. The approved face artwork stays intact while a bounded colour-and-normal transition reconciles the neck. The lesson: painted features must match the underlying geometry; another face stamp is not a substitute for that fit.
• Real rig motion, not a moving capsule. The README includes a complete 1.5-second walking cycle sampled at 24 actual UE animation poses, with the head and sword attached. The GIF contains no generated or interpolated motion.
• Repeatable capture and packaging tools. A JSON recipe stages the character in the existing workshop. The packager checks image hashes, frame ordering, animation times, moving foot bones, attachment errors, and the final GIF's frame count and duration. Nine focused regression tests pass.

What is verified—and what is not

The local packaged character passed 18 programmatic checks covering possession, locomotion, collision, jump/landing, attachments and native mesh budgets. Body, head and sword total 36,437 vertices / 54,027 triangles at LOD0.

The promotional shots are real, unretouched editor captures. A small temporary portrait fill makes the character readable; the saved scene lighting and source assets were not changed. The GIF demonstrates evaluated animation, not frame-rate performance or recorded keyboard-controlled gameplay.

There is still a visible collar edge in close-up, and final neck appearance remains pending human approval. The head is rigidly attached: facial animation and cloth simulation are not implemented. This milestone is not a blanket production-ready certification.

What you can get today

This release publishes source, showcase media, recipes and capture tooling at commit 0cc39d8. It does not include the generated workshop UE project, character source assets, model weights, or a downloadable packaged game. Re-running these specific captures requires the existing local workshop project and verified assets; a fresh clone alone cannot recreate this scene.

• See the updated README
• Read the capture instructions and prerequisites
• Inspect the media provenance

Thanks for following the project. The goal remains the same: turn the hard-won fixes into repeatable tools, and show what actually works—not just the most flattering frame.