KGX export: literal newlines in node descriptions make the TSV mean two things #723

Found by claw's shared KGX checker (culturebotai-claw#273), run against this
repository's own output/kgx/.

What

output/kgx/nodes.tsv is 1,007 rows to anything that splits on newlines and
998 to Python's csv module. Nine node description fields contain a literal
newline — some wrapped in quotes, some not — so the two most obvious readers
disagree about the file.

line 63: CommunityMech:000031  biolink:OrganismalEntity  Geobacter-Clostridium…
line 64: MECHANISM CONTESTED - do not read this record as establishing…   ← same record

line 359: ENVO:00000003  biolink:EnvironmentalFeature  mine tailing  "Abandoned…
line 360: "  infores:communitymech                                      ← same record

Why it matters beyond tidiness

The split shifts columns. provided_by lands in the category position, which
is why a Biolink audit of this file reports infores:communitymech as a node
category on nine nodes (culturebotai-claw#275). There is nothing wrong with the
vocabulary — every biolink: CURIE this repository emits is genuinely in the
Biolink model. The nine are an artefact of the serialisation.

Anything downstream that reads these as TSV rows — cut, awk, pandas with
quoting=QUOTE_NONE, a KGX merge — sees nine malformed records and, depending on
the reader, either drops them or mis-attributes their fields.

Fix

Escape or strip newlines in text fields when writing, or quote consistently.
src/communitymech/export/kgx_export.py writes these; the simplest correct
change is to replace \r and \n with a space (or \\n) in description,
name and supporting_text before writing.

Worth adding a check to validate_kgx.py too, or switching it to the shared
checker:

uv run kg-microbe-kgx check --mech communitymech

which reports exactly this, and which found it here.

Not this repository's problem, but adjacent

The vendored app/kgm/merged-kg_edges.tsv (kg-microbe's merged graph, 1 GB) has
a worse version: 190,506 bare carriage returns and zero CRLF, one of them
inside the header after a duplicated agent_type column. awk reads 11 columns
there; Python reads 7 and silently misaligns everything after it. Raised in
culturebotai-claw#275 — flagging it here only because the copy lives in this
repository.