Problem

PERF-9520 deleted the mongo-perf variants, which held the only collation performance tests in sys-perf:

• insert_read_commands: Insert.IndexedStringsSimpleCollation, Insert.IndexedStringsNonSimpleCollation
• query_read_commands: Queries.StringUnindexedInPredWith(Non)SimpleCollation and the Large variants
• agg-query-comparison_read_commands: the Aggregation.StringUnindexed...Collation counterparts

None of the Locust replacement tickets linked from PERF-9520 covers them. On DSI master the only remaining collation use is the index on foreignLookupFieldDILJ in workloads/query-execution/lookup for insertion, but never be used in measured query.

While testing SERVER-132620, I found that neither QHVW nor HVW exercises collation.

Why it matters

A non-simple collation sends every string comparison through ICU, so extra comparison work shows up much larger in collation tests. They caught:

• BF-31625: 38% regression in Aggregation.StringUnindexedLargeInPredWithNonSimpleCollation from SERVER-84274, fixed by SERVER-85775.
• BF-9690: 50% drop on the large $in microbenchmarks, collation variants included, fixed by SERVER-35851.

Collation also runs on the vendored ICU 57.1, which has a known correctness bug in sort key generation (SERVER-132620).

Proposal

Create a collation performance test in Locust and add it to the Query HVW context. It should cover:

• $in with a non-simple collation, small and large lists
• inserts into an index with a non-simple collation
• sort under a non-simple collation
• one case with strings longer than 250 bytes, which take the heap-buffer path in getComparisonKey after the SERVER-132620 fix; no previous test covered it

The ICU key function itself is covered separately by collator_interface_icu_bm in benchmarks_query (added in SERVER-132620).