fix(build): preserve leaked biomedical category instead of discarding it

[wonhurk]

fix(build): preserve leaked biomedical category instead of discarding it

Problem

Biomedical semantic subagents are instructed (in the /graphify extraction
rules) to record a node's entity category — drug / gene / disease /
protein / technology / peptide / institution / company — in the
node's entry_type attribute. file_type is a separate axis with a small
closed set of valid values (code, document, paper, image, rationale,
concept).

In practice, models frequently also write the category into file_type,
e.g. a patent-record node comes out as file_type: "technology". That is not a
valid file_type. The current canonicalizer in build_from_json handles this by
collapsing any unknown file_type to concept through _FILE_TYPE_SYNONYMS:

if ft and ft not in {"code", "document", "paper", "image", "rationale", "concept"}:
    node["file_type"] = _FILE_TYPE_SYNONYMS.get(ft, "concept")

This silences the invalid-file_type validator warning, but it has two downsides
for biomedical corpora:

1. The category is silently discarded. technology → concept; peptide,
   drug, gene, etc. aren't in the synonym map at all and also fall to
   concept. The information the model extracted is lost.
2. The node is mislabeled as a generic concept document rather than the
   document (patent/paper record) it actually is.

Real-world example (patent records): nodes like mitocatch, phlip,
bispecific-nanobody, erythrocyte-mito-capsule came through with the category
duplicated into file_type — and in the observed data these nodes already
carried the correct entry_type, so the category was recoverable but was
being thrown away.

Fix

When a node's file_type is a leaked biomedical category, recover it into
entry_type (without overwriting an existing one) and coerce file_type to
document:

if ft and ft not in VALID_FILE_TYPES:
    if ft in _BIOMEDICAL_CATEGORIES and (
        node.get("entry_type") or ft not in _FILE_TYPE_SYNONYMS
    ):
        node.setdefault("entry_type", ft)
        node["file_type"] = "document"
    else:
        node["file_type"] = _FILE_TYPE_SYNONYMS.get(ft, "concept")

Ambiguity handling for technology

technology is intentionally also a code-corpus synonym (→ concept), so it
is ambiguous. To avoid changing behavior for non-biomedical corpora, a bare
technology file_type with no biomedical signal keeps the existing
concept mapping; it is only recovered as biomedical when the node already
shows a biomedical signal (an existing entry_type). The condition
ft not in _FILE_TYPE_SYNONYMS means the biomedical-exclusive categories
(peptide, drug, gene, disease, protein, institution, company) are
always recovered, since none of them appear in the synonym map.

Behavior matrix

input file_type	input entry_type	result file_type	result entry_type
peptide	—	document	peptide
technology	technology	document	technology
technology	—	concept (unchanged)	—
gizmo (junk)	—	concept (unchanged)	—
paper (valid)	—	paper (unchanged)	—

No invalid-file_type validator warning is emitted in any case.

Tests

Adds three tests to tests/test_build.py:

• test_leaked_biomedical_category_recovered_into_entry_type
• test_ambiguous_technology_without_signal_stays_code_synonym
• test_unknown_invalid_file_type_still_falls_back_to_concept

All existing test_build.py tests continue to pass.

Scope

Both build_from_json and the multi-extraction build() (which delegates to
build_from_json) are covered by the single change. No public API changes.