Originally created by: 5dive-bot
run-harnesses.sh:889 printed "At least one is marked WRONG (>= N% and >= Ns under): that is not runner variance." as a flat sentence, and :912 turned it into exit 5. It was the last unmeasured "not variance" claim in the file, and it is false.
269e3e7: exit 5, then GREEN on a re-run with no code change. Same sha, same corpus, same claims — the only variable was the runner. 120 harnesses, 0 failing, 500s against a 1320s budget (37%), cross_runner=off.8d661a8: corpus_tier_budget_unit.sh claimed 41.4s and drew 62.1s (+50%) on a branch whose 3-dot diff does not touch it and whose header byte-matches origin/main. It sits on the threshold, so it flips on runner speed alone.This is DIVE-2829's defect one axis over: a one-box sample asserting a cause it has no instrument to separate. Cost: froze the v0.19.14 release cut (~40 min), held PR [#565] and PR [#554], and cost three agents a differential each — because release-cut.yml refuses on any red without reading which red it is.
A WRONG header is still graded, still printed in full with its replacement line. It no longer exit 5s unless --drift-fatal=required. The thresholds in tests/lib/tier.sh are byte-identical — raising them would hide the signal instead of confirming it, which the row explicitly ruled out.
The sentence now prints WHAT IS MEASURED / WHAT IS NOT EXCLUDED, and names why the calibration probe cannot substitute for the missing confirmation: on [#565] it read 101729us/iter against a 119000 baseline — FASTER than baseline — while the workload read +47-73% slow, because the probe prices CPU iterations and these harnesses are priced by process spawn and IO. A probe that samples a different RESOURCE than the workload can neither clear it nor convict it.
DIVE-2829's --cross-runner is deliberately NOT reused. It confirms a second box found the corpus TOTAL over cap, which says nothing about any one header. Reusing it would widen a control to mean something it never measured. This axis has no confirmation instrument and the code says so out loud.
Signal preserved off the PR gate: new report fields header_drift_wrong and drift_fatal_policy — the axis on which a second runner can exclude the slow box, which no single job can. drift_fatal_policy also means a later reader need not date the change to interpret an old report.
All 6 callers (full-sweep.yml x2, unit-tests.yml x4) take any non-zero as a step failure and none distinguishes rc 5, so the effect is precise and intended: test, test-installed-host, both -confirm jobs, full-pristine and full-installed-host stop redding on a drift-WRONG header. None passes --drift-fatal, so the required arm is wired to no caller — stated in the code. Report consumers are append-safe: tier-cal-window.sh looks up named fields with grep -m1 '^# <name>=', and tier-cal-harvest.sh parses harness-budget log patterns the two new lines cannot match.
tests/corpus_tier_budget_unit.sh 135 passed / 0 failed (7 new arms). Touched set also green: full_sweep_pr_trigger 5/0, fold_changelog_fragments 26/0, prose_file_flags 21/0, task_board_write_fence 14/0. bash -n on all three edited files.
One new arm asserts the absence of not runner variance — the old arm required that string, so the test had been holding the false claim in place, and the cheapest revert of "stop asserting X" is to re-add X.
Not a provisioning-path change, so no smoke-gate question. Compiled to community/wiki/an-absolute-time-budget-on-a-variable-platform-measures-the-runner.md (§ RESOLVED 2026-08-10) plus its index line.
Closes DIVE-3163.
🤖 Generated with Claude Code
Ticket changed by: lodar