Story: Harden the gates the component clean standard rests on

A component clean-up passes or fails on what a gate can see. The ores.service clean-up took three verification rounds, and every round failed on something no gate could see: a generated header the drift gate does not cover, a test case the compiler never received, and a record whose numbers went stale when trunk moved. Each round cost a full review cycle, and each found a defect the previous round had declared clean.

This story turns each of those manual checks into a gate. A gate is the only form of the check that survives the reviewer: it runs on every pull request, it runs on a tree nobody measured by hand, and it fails in the same way for the next component as it did for ores.service.

Three classes, one task each:

• Generated output no gate sees. The drift gate regenerates the components the registry lists, and 4,892 of the 5,829 files carrying the generated marker are outside that list. A rename lands in one component's model and writes into another component's tree, where no gate looks.
• A test case that never reaches the compiler. A TEST_CASE inside a false conditional does not exist. The file compiles, the link succeeds, ctest reports green, and only a case count nothing compares against betrays it.
• A record whose numbers go stale on rebase. Every number in a clean-up record is measured at a head that trunk then moves under.

• Each of the three classes is caught by a check that runs on a pull request, with no step a human performs by hand.
• Each new check has a test that fails when the check stops finding anything, so a gate cannot pass every tree by finding nothing.
• Every check the standard names as the proof of an item is in the same CI job the item's story relies on.

• Ban the construct rather than detect the instance. The capture this work came from proposed preprocessing each test source with its own flags from compile_commands.json. That is exact, and it cannot be a pull-request gate: the codegen job has no C++ build, and the job that has one runs on a schedule and on tags. A rule that forbids a TEST_CASE inside a conditional block needs no flags, no build and no formatter, and it makes the class impossible instead of reporting it. The cost is six cases in three files, and the change makes those cases more visible than they were, not less.
• A static rule reads the tree the way the compiler does, or it is a rule about the files it happens to open. The first cut of the gate read only *_tests.cpp and treated a backslash-newline as ordinary text. An independent verifier defeated both: #\ followed by if 0 hid a case, and a case in an included fragment was never read at all. The gate now resolves the splices of translation phase 2 for the whole file before it looks for anything, reads every C++ file under a tests/ directory, and fails on a file whose conditionals do not balance, because a guard left open deletes a case in whichever file includes it.
• A recorded number names its command and its head, and the verb that would re-run them is not built. A clean-up record's counts go stale on their own when trunk moves, so the task template carries a Measurements table with a command and a commit per row and the standard says a count alone is not evidence. The capture also proposed a compass task verify that re-runs a record's declared commands. It is not built: records legitimately carry check_component_drift.py --dry-run, whose non-dry-run form regenerates a component in place, and a verb that runs whatever a document declares would eventually run a destructive one where nobody expects to be asked. The route that would earn its place is a whitelist the gate scripts publish themselves.
• Where the rule cannot read a file, it fails instead of guessing. A compiler takes a raw string literal's extent from the spliced text and its content from the source, while the gate removes every splice. A splice at a literal's opening, or inside its content, can therefore move where the literal ends, and the gate then reads string content as code or code as string. Rather than pick one reading, it reads the raw strings of both texts and reports the file where they disagree. A gate whose silence means clean has to be loud about the inputs it cannot model, or its silence stops meaning anything. The first attempt at this claimed the source reading failed safe, and it did not: the claim was false in the direction that matters, which is how a later verification round found a silent green pass over a real deleted case.

• The registry tier for model-free components, which is what would let ores.service and its five siblings into COMPONENTS_UNDER_TEST. Recorded in the drift-gate capture.
• has_permission() returning true for a token that carries no permissions, a fail-open on an authorization check kept as a recorded pre-RBAC compatibility rule.