Bind the files these fixtures declare

[stevevanhooser]

Closes #293.

VH-Lab/DID-python#86 closed the fail-open in check_files step 3, so a document that declares a required file in its file_list and binds nothing in files.file_info is refused where it is created rather than accepted and failed on at read. #292 fixed the product half (element_timeseries.addepoch); this is the test half — the ~20 construction sites that built such a document.

No schema is relaxed. A generic_file genuinely requires generic_file.ext, the same constraint NDI-matlab#977 respected when makeSourceFile moved to fileReference rather than emitting a fileless generic_file. Each site now attaches a file the way tests/matlab_tests/conftest.py::_add_doc_with_file already did, or uses a document class that declares none.

Measured

Both runs on this container with did and vhlab-toolbox-python installed from main, back to back, same command:

	failed	passed	skipped
before	21	5155	552
after	0	5176	552

comm -13 baseline.txt after.txt is empty — nothing fails that was not already failing. 5155 + 21 = 5176 accounts for every test. ruff check and black --check --target-version py310 pass on all eight touched files.

The two sites that were not "attach a file"

test_extract_doc_files_series.py::test_a_document_with_no_files_is_left_alone wants a document with no files, so it now builds a bare base document instead of a demoNDI. Attaching a file would have inverted the test.

test_did_integration.py::TestBundledDefinitions::test_definitions_validate_or_are_listed is the one the issue flagged as different in kind, and it is. The gate validates a blank template, and a blank template legitimately binds nothing — the document that binds the files is the one that gets stored. The files check now runs ahead of the field checks, so seven definitions began reporting DID:Database:ValidationFiles instead of whatever is actually wrong with them.

That is the right answer about a document and the wrong question about a definition. This is the same situation _BLANK_TEMPLATE_IDENTIFIERS already handles for an unfilled dependency ("a blank template legitimately leaves required dependencies empty"), so the gate now binds the declared names before validating rather than discounting the identifier.

Discounting it would have been the lossy choice, because _structural_failure returns on the first error: five of the seven would have come back clean and tripped the test's own newly_fixed assertion, and the field errors they are listed for would have stopped being checked. Binding instead leaves the allowlist completely unchanged and still earning its keep:

definition	listed as	reports now
data/generic_file	ValidationFieldDouble	ValidationFieldDouble
demoNDI	ValidationFieldInteger	ValidationFieldInteger
demoNDISeries	ValidationFieldInteger	ValidationFieldInteger
demoNDISeriesMixed	ValidationFieldInteger	ValidationFieldInteger
mock/demoNDIMock	ValidationFieldInteger	ValidationFieldInteger
data/geneList	(not listed)	validates clean
sorting/SpikeInterfaceSortingOutputs	(not listed)	validates clean

The last two are the corroboration: they were never on the allowlist, they were tripping only the new instance-level files check, and once it is asked the right question they pass. Nothing was structurally wrong with them, and adding all seven to the allowlist under the new code would have recorded that there was.

A file_list name ending in # is a file series, so it is bound under its first member's name (is_filename_match reads the stem plus digits). Placeholder locations are https:// URLs: can_find_one_file does no network I/O and does not pre-check a non-local location, so the gate still touches no filesystem.

Note, not fixed here

tests/test_session_database_rm.py asserts session.database_search(ndi_query("base.id") == doc.id) == [] in four places. ndi_document.id is a property, so these are real comparisons and the assertions do hold — I checked, because did.document.Document.id is a method and the same line against a raw DID document would be vacuously true. Flagging it only because the two types differ here and it is an easy trap for the next person.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WCfsstd8ma954CDDXg4Hm2

---

Generated by Claude Code