The shipped LRA/SOS front door reconstructs over the constructed reals, and the refutation it returns rests on zero carrier axioms

Counts rather than greps: `grep -q` would pass on ONE surviving fixture if the other two regressed or were dropped, and the second test rules out a mixed result rather than inferring it from the first. The example decides the carrier by the module's carrier DECLARATION, because every `CReal.foo` name contains the substring `Real.foo` and a `contains("Real.lt")` test survives the flip unchanged -- one shipped test did exactly that and asserted nothing. The names in this paragraph are the OLD ones on purpose: ADR-0522 renamed the axiomatized package `Real` -> `AxReal` on 2026-08-19 precisely so that no such substring exists, and rewriting the sentence would erase the hazard it records. Two more sites were found that day by the rename alone, both green.

The footprint SIZE is left free (2, 3, 5) because it is the query's own variables and hypotheses and legitimately differs per fixture; what is pinned is that none of it is carrier. Carrier is computed as the footprint minus the `axeyum.reconstruct.` namespace, which is where this route mints every axiom of its own, so the split cannot be gamed by renaming a carrier declaration.

Without this row the fact would still pass if the AxReal route quietly stopped depending on its own axioms, which would mean the comparison had become trivial rather than that CReal had achieved anything. The second test is the explicit negative: no AxReal fixture may report zero.

The equality of the two counts is computed BY THE EXAMPLE, per fixture, over both carriers, and reported as its own verdict line -- a regex backreference was the first attempt and this host's `grep` does not support one, which would have made the row silently unrunnable. `grep -x` on the whole line including the trailing `true` because an unanchored match also accepts the `false` rendering. This row also refutes a claim recorded on 2026-08-18 in F:farkas-refutation-over-constructed-reals -- that a rendered module over CReal would emit the whole constructed development as axioms and defeat the claim. It does not: `write_lean_module_impl` emits every reachable inductive as a real Lean `inductive`, so the development contributes zero axiom lines and the module declares 3, 5 and 2 (the query's own) against 15, 22 and 10 over AxReal.

awk rather than grep: `checked` MUST equal `modules` and `modules` must be nonzero, because the suite SKIPS itself when no `lean` binary is present and a row asserting only `failed=0` would pass on a host that checked nothing -- and this host's `grep` has no backreferences, so the equality cannot be written as a pattern. `END { exit ... }` with `NR > 0` also fails when the LEAN_CROSSCHECK line is absent entirely. Measured 2026-08-18 against lean 4.30.0: 77 of 77, 0 failed. This row is not decoration: the carrier flip initially FAILED it -- 5 of 77 modules rejected with `Unknown constant Int.natAbs`, because the renderer ordered an inductive by its own type while writing its constructors, which mention later definitions, inline. That defect is fixed in the same change and guarded by two kernel unit tests that need no `lean` binary.

Without `--require-axiom-free` the example exits 0 on a negative result, which is exactly the failure mode this ledger keeps finding; the flag is what makes the status depend on the finding. The four lines are counted, not grepped, so a dropped verdict is a failure rather than a quieter pass, and matched with `grep -x` including the trailing `true` so the `false` rendering of the same sentence cannot satisfy it.