The cleaner treats a prompt as prose or as a comma-separated tag list.
That is right for tags and wrong for the few tokens that are ADDRESSES
rather than words: `embedding:name` and `<lora:name:1.0>` resolve to a
file, so one edited character changes what the sampler loads, or loads
nothing.
Two rules were provably destructive:
embedding:bad_prompt_ -> embedding:bad_prompt
embedding:badhandv4, embedding:badhandv5 -> embedding:badhandv4
`_` and `-` are in the stripped punctuation set, and `bad_prompt_` is a
real published shape. Versioned names sit far above the hardcoded 0.9
fuzzy threshold while naming different files. Both failures are silent:
ComfyUI resolves a missing embedding to nothing, and the dropped tag
only shows up in a generic "N removed" count.
engine/cleaner/atoms.py marks both forms -- LoRA included, since it has
the identical failure mode and excluding it would have been an
arbitrary line. Punctuation stripping skips atoms; fuzzy dedupe falls
back to exact comparison when either side is one.
This widens the cleaner's remit from "text" to "knows a little prompt
syntax", which the findings named as the cost. The alternative was a
rule that renames a file reference without saying so.
Narrow, and the narrowness is tested: ordinary tags still get trimmed,
ordinary near-duplicates still merge, exact repeats of an embedding
still collapse, and text mode still trims a leading edge -- it only
holds back the trailing trim when the prompt ENDS in an atom.
DumiFlex
dafb7d879c