Originally created by: 5dive-bot
Core CI shards go 6 β 10 so a slow runner can no longer push one over the 300s cap.
Why. Three consecutive main runs (44436729689, 44437722568, 44440277989) read core shards of 213-330s, and the merge queue ejected [#1716] and [#1718] with every harness passing. The weights table (DIVE-8748, a day old) reproduces all three partitions 709 of 709, so this isn't a stale table. The shards that go over are the ones that landed on a slow runner: each shard ran at 0.77x to 1.21x its table price.
Sizing. Each split is planned with the shipped table, and its worst shard is costed at the slowest runner measured (1.21x):
| shards | pristine | installed-host |
|---|---|---|
| 7 | 272s (91%) | 286s (95%) |
| 8 | 238s (79%) | 250s (83%) |
| 9 | 211s (70%) | 222s (74%) |
| 10 | 190s (63%) | 200s (67%) |
At each run's own per-harness prices, ten shards give a worst shard of 178s / 187s (62%). The corpus grows about 1.1s per shard per day at ten shards, which leaves roughly six weeks before the slowest runner reaches 85%. Eight shards would pass today and go back over within days.
Changed. Both core matrices, both acp union-grader report lists, CORE_SHARDS 6 β 10 in tests/corpus_tier_budget_unit.sh (arm 98, with the decision recorded), the workflow header, and a changelog fragment. The 300s cap, the calibration and the weights table are untouched.
Checked. bash tests/corpus_tier_budget_unit.sh: 193 passed, 0 failed. Both matrices parse as [1..10].
Not done. Paydown. The five heaviest harnesses add up to only ~93s per environment.
π€ Generated with Claude Code
Ticket changed by: lodar