[stevevanhooser]

Closes #101.

Problem

Query("", "depends_on", name, value) resolves to hasanysubfield_exact_string on depends_on. That operation had no SQL form, so every depends_on search loaded and JSON-parsed every document in the branch, then tested each one in Python.

NDI uses depends_on constantly: "all documents that depend on this element / epoch / probe".

Change

depends_on is now matched in SQL against meta.depends_on, which doc2sql stores as name1,value1;name2,value2;.

• Match rule. An entry (name, value) is present exactly when ';' || meta.depends_on contains ;name,value;. For the * name, the stored string must contain ,value;.
• Matcher. instr() is a plain, case-sensitive substring test, like brute force's ==. MATLAB's own SQL for this, LIKE '%name,value;%', is not copied: it ignores case, and it isn't anchored, so id,X; would also match subject_id,X;. The bridge decision_log records this difference.
• Kept on brute force: a query whose name or value is empty, or contains , or ;. Neither case can be expressed exactly in that stored format.
• Order within an AND. A depends_on that has an SQL form now ranks with the ordinary terms, ahead of base.session_id. It's one row per document, tested in SQL.

Also fixed (from #99)

Creating a new database printed three DID: adding search index … (about 0 rows) lines. _create_db_tables doesn't create the indexes, so _ensure_indexes builds all three on every new file. It now builds them silently when doc_data is empty.

Timings

50,000 documents, each with 2 dependencies; main → this PR:

search	main	this PR
depends_on element_id E7 (100 docs)	2.07 s	0.009 s
depends_on * E7	0.79 s	0.007 s
depends_on element_id E7 AND session	1.19 s	0.009 s

Tests

New file tests/test_depends_on_search.py. It compares the SQL path with brute force, for both depends_on and ~depends_on, on documents built to trip it up:

• id vs subject_id;
• E1 vs xE1 / E10;
• case differences;
• enumerated names;
• the * name;
• empty values;
• documents with no dependencies.

It also checks that no document is loaded, and that what SQL can't represent still uses brute force.

Other test changes:

• test_search_indexes: a new database is created without printing anything.
• test_query_correctness: the depends_on SQL uses no LIKE wildcards.

Local results:

• black --check and ruff check are clean.
• pytest passes: 823 tests.
• Bridge coverage, all six checks against DID-matlab main: 0 problems.

---

Generated by Claude Code