The cross self-host proof runs a different configuration than the native one

• Type: bug (undisclosed qualifier on a load-bearing claim) — Track A owns the reason; the docs half is already handled.
• Found: 2026-08-30 by frankD, which verified the claim rather than accepting it — reading Makefile:13809's cross-bootstrap rules to confirm the cross proof is a real proof (cross-compile compiler.pas for the target, run that binary under QEMU to compile it again, cmp), and finding the configuration difference on the way.

The two configurations

So "all four self-host byte-identical" reads as one gate and is two. Both are real; they are not the same claim.

What Track A needs to supply

Why the managed runtime is required for the cross bootstrap. frankD deliberately stated the difference without explaining it — "asserting a reason would be inventing one" — which is the right call and the reason this ticket exists rather than a docs edit that guesses. The candidate readings differ a lot in what they imply:

• a deliberate, permanent property of cross targets (then it is documentation, and the claim should say so on both surfaces); or
• a workaround for an unmanaged-string defect on non-x86-64 targets (then it is a bug with a -dPXX_MANAGED_STRING shaped scar on it, and the cross claim is weaker than it reads); or
• historical, no longer required, never re-tested. Cheapest to check and the most likely to be true of a required flag nobody has questioned — cf. face 181, a guard already hardened once is the one whose stated reason nobody re-reads.

A one-sentence answer closes this. frankD has offered to place it in docs/**.

Why it matters beyond the wording

Same failure mode as the -O-scope omission found in the same audit, one axis over: a true claim whose qualifier lives in a Makefile comment no reader will ever see. The claims-discipline section of CLAUDE.md is about not conflating two different byte-identical claims; this is a third distinction inside one of them.

And it bears on the self-host gate's own authority. CLAUDE.md already narrows that gate once — "evidence the compiler compiles itself at one optimisation level, not that it compiles itself" — after a -O0-only self-compile failure passed the entire gate on 2026-08-19. This is the second narrowing, on the axis of build flags rather than optimisation level, and it was likewise found by someone looking at something else.