Originally created by: lodar
Fixes main red: unit-tests core-pristine-confirm (6) + test-confirm @6e1e2516 (run 44421508722), exit 4 with 119/119 harnesses green. Budget red, not a test failure.
Cause: not [#1712] (it touches model names and two tests on other shards). Shard 6 was already 330s on its first box at the previous green sha a96434f5 (run 44298553604); that run passed only because the confirm box drew 296s. This time the confirm box drew 331s against a 309s effective cap. The 2026-11-13 weights table had drifted: the planner loaded shards 6 and 1 (327s / 299s) and left shards 2 and 4 near 220s.
Fix: ran scripts/refresh-harness-weights.sh (core run 44298553604, full run 44298553674) and regenerated tests/lib/harness-weights.tsv. Added changelog.d/DIVE-8749.md. No cap, shard count or harness changed.
Check: tier.sh list core | tier.sh shard 6 <env> <tsv> (709 core harnesses, 0 unpriced), each plan priced at the red run's own per-file times from its shard report artifacts:
Residual: pristine core is 1598s over 6 shards, about 89% of capacity on average, so this buys days, not weeks. Installed-host passes only through runner calibration. The next heavy harness reds again whatever the table says; that is a merge/retire/shard decision this PR does not make.
🤖 Generated with Claude Code
Originally posted by: lodar
Superseded by [#1718] (DIVE-8748, graded PASS at fbabb127): same harness-weights regen. Closing to avoid a conflict race.
Related
Tickets:
#1718Ticket changed by: lodar