The read-cost gate fails on how loaded the machine is, which has held trunk CI red and the docs site undeployed since 2026-09-09

Trunk CI is green again, and the docs site deploys again. Both have been down since 2026-09-09 11:20: every Rewrite reactor CI run on claude/graphitron-rewrite since then has failed, so docs-build and docs-deploy, which are gated on the build job, have been skipped rather than run, and graphitron.sikt.no has not been republished for two days. Nothing is wrong with the documentation. The docs jobs are collateral, and anyone reading the run list sees docs not happening and infers a docs problem, which is the second thing this item fixes.

The failing assertion is DerivedReadCostTest.theCellsThatCouldNotBeComparedAreExactlyTheOnesRecorded. That gate prices each derived relation with a registration on and off and compares rows visited; a cell whose unregistered side does not answer inside a relative wall-clock budget is recorded as unmeasurable instead of compared, and KNOWN_EXHAUSTED pins the recorded set at empty. On the CI runner two cells exhaust it, intent_carrier_data_field|diagnostic and intent_carrier_data_field|intent_field_unlowerable_ordering, so the pin fails. Locally the same class is green.

The defect is in the premise, not in the number, and the pin’s own javadoc states the premise it has stopped satisfying: BUDGET_FLOOR_MILLIS "is set clear of the slowest of them so that which cells appear here is a fact about the code and not about how loaded the machine was". On a shared GitHub runner it is not clear of the slowest, so the gate now reports a machine fact as a code fact. Raising the floor until CI passes is the move to refuse: it weakens the one gate whose job is noticing a reader getting dearer, and it would have to be raised again on the next slower runner.

That this is environmental rather than a regression is established rather than assumed. The identical assertion fails with the identical two cells on 34349521966, whose commit is a roadmap-only diff touching one markdown file, so no code change can account for it. It also long predates R939, whose registration landed 2026-09-10 13:20 and which moved this test’s other pins without touching these cells.

Two things this item owes beyond the fix. Whatever replaces the wall-clock cut has to keep the arm that aCellThatCannotAnswerIsRecordedRatherThanFailed exercises, because a genuinely non-terminating unregistered side must still pass rather than fail: non-termination is the strongest form of "materializing did not make this worse". And the docs jobs should say why they did not run, since a skipped deploy currently reads the same as a deploy nobody asked for.