Activate the MONDO close-match guard in disease glom (with build-vs-build clique-diff tool)

[gaurav]

Important

Superseded by #888. The MONDO close-match guard fix has been split onto a clean,
main-based branch in #888 (the same x[1]→x[2] change applied directly to
build_compendium, plus a focused regression test), and the build-vs-build clique-diff tool
already merged separately via #885. If #888 merges first, this guard change drops out of this
PR automatically once main is merged in — leaving #883 with nothing unique to contribute.
Prefer reviewing #888 for the guard fix in isolation; this PR is kept only as the home of the
full add-mpo-base before/after impact analysis under
docs/pipelines/diseasephenotype/mondo-close-match-guard/.

Summary

Fixes a latent bug in how the disease/phenotype build reads the MONDO_close concord, which
left glom's close-match guard a silent no-op, and adds a reusable build-vs-build clique-diff
tool plus the before/after impact analysis that this change requires.

This is the dedicated follow-up promised in #790, where the same code was touched only enough to
stop the build crashing (validating the real 3-column MONDO_close format while preserving
the dormant x[1] behaviour). Activating the guard changes disease clique merging broadly and is
orthogonal to adding MP, so it lives here with its own evidence.

Note

Based on add-mpo (#790), not main — it builds on that PR's compute_cliques_for_impact_report
refactor. Merge #790 first; this PR will retarget to main automatically.

The bug

compute_cliques_for_impact_report() read MONDO_close (a 3-column subject predicate object
concord) with close_mondos[x[0]].add(x[1]) — keying the MONDO subject to column 2, the
predicate (oio:closeMatch) instead of column 3, the close-match object. No clique ever
contains the literal string oio:closeMatch, so glom's close= guard (if cd in newset) never
matched. The guard has therefore never fired, on main and before it; close (but not exact)
matches were free to collapse into exact cliques.

The fix is one line of intent: key on x[2], matching the (stuff[0], stuff[2]) parsing every
other concord in that function already uses.

Impact (why this needs SME review)

Measured on a full local disease build (Biolink 4.4.3), same intermediate ids/concords,
differing only in x[1] vs x[2]. glom is deterministic — two x[1] builds were clique-identical
— so every difference is signal, not run-to-run noise.

• 1,219 identifiers, all MEDDRA, are dropped from Disease.txt. They were present only because
  the dormant guard let them merge into a MONDO disease clique; correctly kept out, they have no
  independent disease typing in this pipeline and leave the compendium entirely.
• 1,191 of 365,465 Disease cliques lose ≥1 member; net Disease clique count +20, PhenotypicFeature +13.
• PhenotypicFeature.txt is essentially unaffected (the close pairs are MONDO-subject).

Dropping 1,219 MEDDRA codes is correct if they are genuinely distinct concepts (the guard's
premise), but a coverage loss if anything relies on that MEDDRA → MONDO normalization. Per the
project rule against dropping valid identifiers without good reason, this is an SME call — the
evidence is in docs/pipelines/diseasephenotype/mondo-close-match-guard/ (README.md,
clique-diff.csv, dropped-members.csv, summary.json).

New tool: babel-clique-diff

A reusable build-vs-build clique diff (tools/clique_diff) that compares the finished compendia of
two builds and reports split/merged/lost cliques and, crucially, dropped CURIEs.

It is deliberately distinct from source-impact-report: that answers "what does adding source
X do?" by re-glomming intermediates with vs. without one source; this answers "how did the cliques
change between build A and build B?" given the same inputs but different code/config/data.
Because it works on finished compendia rather than glom state, it can also compare a local build
against a published stars.renci.org build with no re-glom — useful for validating any glom-logic
change (close-match, unique_prefixes, overuse filtering) or as a release regression check.

Test plan

• uv run ruff check / ruff format --check — clean.
• uv run rumdl check . — clean.
• uv run pytest -m unit -q — green (243 passed), including new tests/tools/test_clique_diff_tool.py.

Out of scope

• The MP source-impact report on Add MP (Mammalian Phenotype Ontology) as a disease/phenotype source #790 reflects x[1] (the production behaviour of that PR) and is
  not regenerated here.
• Deciding whether to keep the guard active — that is the SME question this PR's evidence exists to inform.

🤖 Generated with Claude Code

[gaurav]

Heads up: the babel-clique-diff tool commit from this PR has been extracted into its own standalone PR off main (#885) so it can land independently of the MONDO close-match-guard change here. Once #885 merges, please rebase this PR to drop its now-duplicated babel-clique-diff commit. The new "MP kept disjoint from HP" PR (#886) depends on #885.

[gaurav]

Closing: everything in this PR now lives either in main or in #888.

Verified against main (9058e63) and #888 (fix-mondo-close-guard):

• babel-clique-diff tool (tools/clique_diff/, pyproject.toml entry point, docs/tools/README.md, tests/tools/test_clique_diff_tool.py) — merged via Add babel-clique-diff: build-vs-build compendium clique diff tool #885, in a better shape than here: the diff logic now lives in src/model/compendium_diff.py with src/tools/clique_diff/cli.py as a thin CLI, and the tests moved to tests/tools/clique_diff/test_cli.py.
• The MP base this branch was stacked on (Add MP (Mammalian Phenotype Ontology) as a disease/phenotype source #790, since closed) — the MP source, its docs under docs/sources/MP/, and tests/pipeline/test_mp_pipeline.py are all in main.
• The guard fix itself (close_mondos[x[0]].add(x[1]) → x[2]) — carried by Activate glom's MONDO close-match guard, with concord-reader symmetry tests #888, applied directly to build_compendium on a clean main base, with 164 lines of new concord-reader symmetry tests in tests/createcompendia/test_diseasephenotype.py that this PR did not have.
• The before/after impact analysis (docs/pipelines/diseasephenotype/mondo-close-match-guard/) — superseded by docs/sources/MONDO/mondo-close-match-guard/ on Activate glom's MONDO close-match guard, with concord-reader symmetry tests #888, which is a strict superset: same clique diff (1,218 dropped MEDDRA identifiers, matching the ~1,219 reported here), plus the split-cliques.csv analysis of the 37 cliques that structurally split — the changes that actually need SME review — and the explanation of why glom() ejects a legitimate non-MEDDRA identifier as collateral damage when the guard fires.

The one file with no direct counterpart on #888 is dropped-members.csv, but it is redundant: all 1,218 dropped CURIEs are fully enumerated (no truncation) in the example_members column of #888’s clique-diff.csv, keyed to the MONDO clique each left.

Nothing needs cherry-picking. The guard fix continues on #888.