Training only one TE creates a DB that I can't extract from and when loading Automatic1111 and Comfy UI reports errors.

It reports the errors in the TE I did not train and I can't extract to LoCON, or LoRA as the TE dictionary is bad with invalid shapes.

No files, tests, or entry points are identified. Start by reproducing the single-text-encoder training flow and the reported extraction and loading errors in Automatic1111 and Comfy UI, then compare the saved text-encoder dictionary and its shapes. Done means the untrained text encoder no longer produces invalid shapes and the resulting database can be extracted to LoCON or LoRA and loaded successfully.

Written by the indexing model from the issue text.