[merynz]

Why

Master source byte recovery and a per-head truth matrix are necessary but insufficient. We need real imported-source observations at scale before saying which candidates can train which head. Keep source-original, imported observations, semantic and Graphics distinct.

What

• Deterministic SHA and file-type diverse FIT pilot from private recovery receipts: up to 8 sources per 8 separated recovery shards (up to 64, GLB/FBX/glTF).
• Blender 4.5.14 pinned binary, isolated per-source processes with --disable-autoexec and original-byte exact source SHA verify.
• Full gzip evidence per source: native source scene nodes, imported mesh topology/open surfaces, skin sample count, material+missing texture dependencies, armatures, bone/constraint/controller/driver graph, NLA and F-curve inventory.
• Brief per-asset JSON and per-shard batch status, separately enumerating import errors rather than false PASS. No copyrighted source file or private source URL enters GitHub; outputs private Drive only.
• All six head routes marked as observation/candidate only; 2D Graphics target requires separately verified stylized art. Original rights and class unknown.

State

Synthetic tests passing. Remote source import wave run https://github.com/merynz/RealSaS-OPT/actions/runs/42387604493; counts will be reconciled only after actual receipts finish. No training or merge authorized.

Risk

Blender source-compatibility may expose importer exceptions. The entire point is to detect and itemize these before training; no asset-specific heuristics or silent fallback.