Originally created by: lodar
task_core_unit (28/7) and task_verifier_rail_unit (17/6) were red on the control-plane host and GREEN in CI at the same commit. It presented as host state — the DIVE-1919 class test-installed-host exists to catch — and it was not. Seeding a runner would never have reproduced it.
FIVE_VERIFY_DEFAULT=0 is in the caller's environment. The DIVE-969 verifier-by-default rail reads ${FIVE_VERIFY_DEFAULT:-1}, so the whole mechanism goes inert — not a wrong value, no mechanism at all (verifyDefaulted:false, verifySkipped:false, verifier:""). Reproduced exactly: with that one export the two harnesses give 28/7 and 17/6, same arm names.
lodar set it for sixteen agents on 2026-07-29 00:49 via the unit's EnvironmentFile=-/var/lib/5dive/agents.d/%i.env, and every file carries the decision and its own revert instruction. Tasks filed without a rail since then are policy-conformant, not damage — there is no backlog of broken rows to audit.
That is exactly why the harnesses are the only thing that can move. The configuration is not an accident awaiting cleanup; it is correct, intentional and permanent. A harness asserting what DIVE-969 does by default while reading that default from the environment is grading fleet policy instead of the code.
DIVE-2211 one axis over. That change pinned WHICH TREE a harness grades, because "21 passed" was a claim about whatever was on disk. The same sentence was still true of the ENVIRONMENT: src reads 14 caller-overridable FIVE_* knobs.
tests/lib/env_isolation.sh unsets the whole FIVE_* namespace and announces on stderr what it cleared rather than silently fixing the reader up. Silent on a clean environment, so the line means something when it appears.tests/lib/grading_tree.sh — all 235 harnesses already source it near the top before any fixture setup. One seam, no 235-file diff. Verified no harness assigns a FIVE_* variable before that point.FIVE_GATE_REPOS, FIVE_GATE_MAIN_BRANCH, FIVE_GATE_ANCESTRY_SCAN would each silently rewrite what the merge-gate harnesses measure, and knob [#15] gets added by someone who never read the file.:- default that ships in src/, which is by construction what CI runs with. A hardcoded default here would be a second copy of a constant that lives in src and would drift.tests/env_isolation_unit.sh — 13 arms, mutation-graded five ways with the measured results in its header:
| mutation | result |
|---|---|
| delete the source line from the seam | 7/3 — T5, T7, T8; T7's output shows the rail harness back at 17/6 |
| unset without printing | 8/2 — both T2 arms |
| print unconditionally | 9/1 — T3 |
filter also matches a non-FIVE_ var |
9/1 — T4 |
widen filter to grep '' |
VOID, recorded so nobody repeats it |
T6 anchors T5 (a variable that was never going to be set proves nothing when found unset). T7 reproduces the incident end-to-end. T9 exists because the void mutation caught the helper announcing the same readonly knob as both stuck and cleared — a report of work that did not happen, which is not allowed to live in the reporter for that defect class; it also revealed that unset on a readonly var would abort a set -e harness, so every step is now individually non-fatal.
Full 236-harness suite green on a committed clean tree, tree unchanged after the run.
The instrument misled first. A /proc/<pid>/environ sweep found the knob in one session and that read as a stray export by one operator. A live environ reflects the env file as of that session's last exec, so the sweep measured restart order, not configuration — agents that hadn't restarted since 00:49 still showed a pre-policy environ. Read the config source, not the running processes.
Ticket changed by: lodar