Markdown export silently overwrites when two sources differ only by extension

[charitarthchugh]

get_markdown_path() in src/paperscale/pipeline.py:691 builds the output filename with:

md_filename = os.path.splitext(os.path.basename(relative_path))[0] + ".md"

It drops the source extension. paperscale accepts both PDFs and images (--pdfs takes
"Local PDF/image paths, a local glob ... or a .txt file listing paths"), so two source files
in one directory that differ only by extension collide onto a single markdown file:

corpus/case.pdf   ->  workspace/markdown/corpus/case.md
corpus/case.tiff  ->  workspace/markdown/corpus/case.md   # overwrites, silently

Whichever document is exported second wins. Nothing warns, and the loss is invisible unless
you count the output files against the input list.

Related precedent: evaluation/runs.py:45 already treats a duplicate Source-File inside one
run as fatal (DuplicateSourceFileError) on exactly the grounds that an ambiguous document key
is not recoverable. The export path has no equivalent guard.

Found while resolving #22. That ticket chose extension-replacement for the embed output too, so
the same collision family will exist on the .npz side and both should be fixed by the same
rule — see #22's resolution, which makes the path-hash the tiebreak.

[charitarthchugh]

Surfaced while resolving How resume knows a document is done — the same function has two further collision classes beyond the extension one this issue reports.

get_markdown_path at src/paperscale/pipeline.py:686-689:

parts = relative_path.split("/")
safe_parts = [p for p in parts if p and p != ".."]
relative_path = "/".join(safe_parts)

1. .. is dropped, not resolved. /a/b/../c.pdf and /a/b/c.pdf both derive a/b/c.pdf, so one export silently overwrites the other.
2. . is kept. /a/./b.pdf derives the string a/./b.pdf while the file lands at a/b.pdf, colliding with /a/b.pdf — and the recorded name does not match where the file actually is.

The fix is to normalize before sanitizing rather than instead of it: os.path.normpath first, then lstrip("/"), then drop remaining .. and empty parts. normpath resolves interior .. and . and leaves leading .. alone, so the path-traversal guard is still needed after it.

Both are pre-existing on main and are not part of the embed map, but the embed design mirrors this rule by standing decision 6, and #28 adopted the normalized version for it. If the export is not fixed the same way, the markdown and vector trees stop matching — which is the property standing decision 6 chose this rule for.

[charitarthchugh]

Decision — add the guard, keep the current names

Settled while resolving #41.

get_markdown_path() is not changed. Instead the export gains a guard: detect two Documents deriving the same markdown path within one export, and fail loudly naming both raw Source-File values, rather than letting the second silently overwrite the first.

Why the guard and not the rename. It matches DuplicateSourceFileError (src/paperscale/evaluation/runs.py:45), the precedent this repo already follows for an ambiguous document key — which is the argument this ticket's own report makes. Renaming every existing markdown file is a larger change than this bug warrants, and markdown output paths are shipped behaviour.

What this knowingly does not fix. --pdfs accepts "Local PDF/image paths", so case.pdf and case.tiff in one directory is a supported input. After the guard it stops the export instead of losing a Document. That is better, and it is not fixed — the corpus still cannot be exported without renaming a source file by hand.

embed went the other way, deliberately. #41 chose extension-append for the embed Document name (case.pdf.npz), which prevents this collision class rather than detecting it. So #22's "mirror the markdown export" rationale no longer describes either side.

Both loose ends — the unfixed class here, and the divergence — are filed as #42: one shared derivation for every output path, with the migration handled openly.

Measured while deciding: the live corpus is 39,905 files, all .pdf, with zero collisions of any class. Nothing is losing data today.