Every change that adds a corpus program also added its CI costs to
tests/harness/corpus-costs.json. The table is sorted by program name and
programs are numbered in sequence, so concurrent changes inserted lines at
the same place and conflicted. A change now writes
tests/harness/corpus-costs/<program>.json ({ "plain": ms, "san": ms }).
The shard loader applies these files over the table, and
scripts/ci-corpus-costs.mjs folds them into the table and deletes them when
it refreshes the costs from CI logs. With no per-program files the table and
the shard assignment are unchanged.
The differential and llvm-differential corpus split their programs across the ten Linux test shards by a hash of
the program's name, so a shard's share of the measured work varied from 565 to 913 s (plain) and 868 to 1358 s
(sanitized). corpusSelect assigns programs longest first to the lightest shard by their measured cost per flavor
(tests/harness/corpus-costs.json: the differential plus llvm-differential milliseconds on a CI runner, medians of
four green main runs), giving 740-741 s and 1098-1099 s. Both harnesses still select identically, so a shard's
compile cache serves both. Programs without a measurement weigh as much as the median one.
scripts/ci-corpus-costs.mjs refreshes the table from downloaded CI shard logs.
Chris Tate