XL-0N-RC: prove the victim is in the measured closure — it is, and the Rc<i64> span-stamping arm is not live on main

[gunbai-bot]

What this is

The XL-0N-RC lane deliverable, in the ordering the brief fixed: positive control before any interpretation. A predecessor lane reported the `Rc` span-stamping defect REFUTED off an A/B over `00_compile.dag`, then retracted it — the victim (`std_checked_arithmetic.rs`) was never in that closure, so both arms agreed at zero for want of a subject.

What was measured (all on current main eced9d2 merged)

1. Subject reached: the import chain `v2.std.integer → std.integer → std.induction → std.checked_arithmetic` puts the victim in the census closure of `src/v2/compiler/03_normalize.dag` (80 emitted files, both Nat authorities present).
2. Instrument discriminates: with `decl_file_realizes_natively` forced false (compiler-side perturbation, reverted), the same probe reds with 2 errors inside the victim and `Nat` rendered structurally. The green arm's zero is a real zero.
3. Key B (decl_file via ident_span on the alias): hits — measured both directions.
4. Key A (bare vs qualified spelling): a two-spelling fixture emits byte-identical code; the un-peeled `authored_name_at` in `field_access_field_is_boxed` is real but has no authorable RED at field grain (`needs_box_wrapping` peels and boxes only recursive carriers), so per DESIGN §4b no check or repair ships against it — the family stays on its existing carrier (`rust_carrier_realizes_as_machine_scalar` note, rung Mitigatable, ceiling: resolved-declaration-identity keying).

Changes

• `docs/plans/self-host-cargo-refusal-root-partition.md`: new §23 recording the subject-present measurement with named instruments (per §6, commands not transcribed numbers where re-derivable), replacing the retracted verdict so the hypothesis is re-read rather than re-run.
• Same doc: the stale citation of `nat_max_two_nat_authorities_note` (deleted by XL-0N residue: refusals carry the BinOp coproduct, the Nat fork gets a stall row, prose leaves the data plane #9794) now cites its true referent, the stall row `gunbc.guarantee_rung_drop two_nat_authorities_stall`. The doc is a `HandAuthoredDocBind` in `doc_graph_roots.dag`, not a main_wet projection — verified before hand-editing.

No compiler, emitter, or resolver code changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VdQphfzTtHwuDqdNyuQQTC

Amendment (same day): the "no authorable RED" verdict is RETRACTED — wrong boundary

The manager's review question ("which boundary is the unauthorable-RED measured against — fixture or corpus?") was decided by execution against the first reading: the RED is authorable as fixture source. A recursive carrier that merely shares the bare spelling Nat, placed in the closure beside a std.nat.Nat field consumer, is accepted by gunbc and refused by cargo with the verbatim claimed diagnostic expected i64, found Rc<i64> — the numeric guard hits while the recursive/shared layer sets are keyed on the bare string and get polluted by the unrelated declaration. The first §23 sentence had measured corpus occupancy, not fixture reachability (reachability_read_as_occupancy). §23 now records the retraction, the executed two-module reproducer, and states the #9813 non-reproduction explanation as a hypothesis with its discriminator rather than an attribution (the whole-corpus attribution boundary has not published). Enrolling the reproducer as an expected-red row and the identity-keyed repair of the layer sets are deliberately follow-up increments, not part of this PR.