BF-46328 traced an ARM64 performance regression to two consecutive PGO profile refreshes on v8.3: the Sep 8 refresh (fe38bd9, SERVER-134346) that the BF caught, and the Aug 31 refresh (aaa6c7c2, SERVER-134047) which had already regressed the same aggregation and expression workloads with no BF filed. Both refreshes changed only the aarch64 profile entries, and the regression is ARM64-only.

Intel performance changes on the same commit are disk and system metrics and belong to BF-46351.

This ticket reverts the ARM64 clang PGO profdata and BOLT fdata (URLs and checksums) to the Aug 24 refresh (18522b34, SERVER-133730), skipping both bad refreshes. The x86 data is unchanged. No GA'd 8.3 build carries the regressed data (r8.3.11 predates both refreshes); 8.3.12 would be the first, which is why this blocks that release.

This is the same process as BF-46248 / SERVER-134811 on 9.0 (aggregation and SBE code has no coverage in the PGO training corpus, so each refresh lays it out blind). That revert is in review on v9.0-staging. The next automated profile refresh will supersede this change.

PR: (will add once opened)