One document-name derivation, shared by every output path

[charitarthchugh]

The problem

paperscale derives an output name from a Record's Source-File in more than one place, and after #41 those places no longer agree.

path	rule	collision behaviour
get_markdown_path() (src/paperscale/pipeline.py:674)	replaces the extension	silently overwrites (#32); a guard is being added, so it will stop loudly instead
embed's .npz Sink (#22, #41)	appends the extension	cannot collide on extension; anything else is fatal at startup
embed's LanceDB document_name (#27)	same as .npz	same

#32 chose the guard over the rename deliberately — renaming every existing markdown file is a larger change than that bug warrants, and a guard matches the DuplicateSourceFileError precedent this repo already follows. That is the right near-term call. It leaves two things unfixed.

1. A mixed corpus still cannot be exported. The guard turns a silent overwrite into a hard stop. --pdfs accepts "Local PDF/image paths", so case.pdf and case.tiff in one directory is a supported input that now fails instead of losing data. Better, not fixed.

2. The two derivations have diverged. #22 chose extension-replacement specifically to mirror get_markdown_path rather than diverge from it. #41 broke that mirror on purpose, because the thing being mirrored is the defect. So the stated reason for a shared rule still stands while nothing actually shares one.

Three collision classes; only one is addressed anywhere

class	example	prevented by extension-append?
extension replacement	case.pdf + case.tiff	yes
sanitization — lstrip("/"), .. and empty components dropped	/a/case.pdf + a/case.pdf	no
tarball collapse	x.tar.gz::doc.pdf + x.tar::doc.pdf	no

Both remaining classes are fatal-at-startup in embed and silent in the markdown export.

What this ticket would do

• One derivation function, used by the markdown export, the .npz Sink and LanceDB.
• A migration decision. Changing the markdown rule renames every existing export. Markdown is regenerable from results/*.jsonl via --export-markdown, so the cost is one command plus clearing the old tree — but it is still a breaking change to shipped output paths and should be announced rather than slipped in.
• One collision policy across all three outputs, covering all three classes.

Not urgent — measured

The live corpus is 39,905 files, all .pdf, with zero collisions of any class. The 49-Document smoke set derives 49 distinct names. Nothing is losing data today. This is debt to pay deliberately, not a fire.

Related

• Markdown export silently overwrites when two sources differ only by extension #32 — the markdown export bug and the guard chosen for it
• The document identity rule: one name, three jobs #22 — the embed name rule and its "mirror the export" rationale
• What happens when two Documents derive the same name #41 — the collision policy settled for embed