BF-46248 traced a performance regression to the Aug 31 automated PGO profile refresh (9061da2, SERVER-134109): the resulting ARM64 code layout costs 8-10% throughput on mixed_workloads_with_scans_locust (every operation type) and ~12% latency on top_twenty_shapes_locust (customer aggregation shapes), across all ARM64 sys-perf-9.0 variants. Indexed reads are unaffected.

Confirmed on 2026-09-10 with a pure A/B: waterfall task activation at the two boundary revisions (orders 1903 and 1904, which differ only by the profile-data commit), rerun on fresh hosts, two runs per side per variant. Every pre/post pair is non-overlapping and the top_ten_queries control is flat to 0.1%.

No GA'd build carries the regressed profile (the r9.0.0 and r9.0.1 tags predate it), but the v9.0 branch does, so the 9.0.2 cut would ship it. This ticket reverts the ARM64 clang PGO profdata and ARM64 BOLT fdata to the Aug 24 refresh (bd692f5, SERVER-133787), which measured clean from Aug 24 to Aug 31. The next automated profile refresh will supersede this.

PR: v9.0-staging: https://github.com/10gen/mongo/pull/62538

Full A/B analysis will be posted on the BF ticket.