Stop conflate_synonyms computing output it then discards

[gaurav]

conflate_synonyms() drives two of Babel's most expensive rules — geneprotein_conflated_synonyms (15,719 s, 98 GB max RSS) and drugchemical_conflated_synonyms (9,949 s) — together about 8.5% of the pipeline's summed rule wall time, both ~99% CPU-bound on one core.

get_logger() defaults to logging.INFO, so logger.debug() output is suppressed — but an f-string argument is still evaluated in full before the call. Four sites were paying for results nobody reads.

The four sites

• :143 — the whole accumulated synonyms_to_conflate structure serialised with json.dumps() to feed one suppressed debug line. That structure is the 98 GB of resident memory.
• :260-261 — every output record serialised twice more for suppressed debug lines, once with indent=2, on top of the json.dumps() that actually writes it.
• :135-140 — every input line recomputed a bl_types set over every synonym accumulated so far for its preferred ID — quadratic in clique size — and used it only in a suppressed debug line. It also rebound synonym inside the read loop, shadowing the loop variable.
• :168, :170, :172-174 — the innermost step-3 loops logged at INFO, one of them formatting an entire synonym list per synonym.

Remaining logs take lazy %s arguments so nothing is formatted unless the level is enabled.

Measured

Over the local anatomy build's synonyms and compendia, with conflation groups of 2, 8 and 32 members: 1.33×, 3.11× and 1.39× faster, output identical every time.

Those inputs are three orders of magnitude smaller than the real gene/protein ones, so they do not predict the production figure — the site that should dominate there is the whole-structure json.dumps, which scales with a structure this data barely populates. Worth re-measuring on a real run.

Tests, and a pre-existing bug left unfixed

This adds the first tests this function has had. Records are copied verbatim from a local anatomy build; only the conflation file is authored, since a conflation pairing anatomy cliques does not occur in a real build (it stands in for GeneProtein.txt, same JSONL shape).

The third test pins a data-loss bug found while writing them, deliberately left unfixed here because this commit is meant to be behaviour-preserving:

ids (synonymconflation.py:87) is a one-shot map object. The outer loop searching for a conflated identifier consumes it, so the inner loop that registers the clique sees only what came after the match. The matching identifier — the one the conflation file names, and the one step 3 looks up — is therefore never registered, so a synonym record keyed on a non-leader identifier of a conflated clique is silently dropped rather than merged.

Verified identical on origin/main, so it is pre-existing. The test asserts the wrong-but-current behaviour with an invert-when-fixed comment, per tests/CLAUDE.md. It needs a tracking issue — I did not file one, and the test comment says so rather than pointing at a number that does not exist.

🤖 Generated with Claude Code