feat(dlv): the 0x0026 bundle consumes the generic proof, and a trader can build one

[cryptskii]

Summary

The 0x0026 cut: the reserve-consumption bundle consumes the generic proof artifact, and a trader can build one the production verifier accepts.

routing advertisement            (locator only)
        ↓
(owner economic position, artifact address)
        ↓
resolve the owner's write-once R_econ root      ← the verifier's own walk
        ↓
fetch + verify EconomicProofArtifactV1          ← against coordinates the ARM derived
        ↓
select the exact vault reserve leaves           ← this vault, this generation
        ↓
owner authority from the authenticated presentation
        ↓
ReserveConsumptionEvidenceV1  →  the REAL 0x0026 verifier  →  build_write_set

One proof source, not two

The bundle carried its own reserve_a_state / reserve_a_siblings / reserve_b_state / reserve_b_siblings — a second representation of exactly what the artifact proves. Those are now reserved, replaced by economic_proof_addr. Two representations of one fact can disagree; one cannot.

The arm fetches the artifact by the same inner content identity every object in this DAG uses, re-hashes it to the named address, verifies it against the owner, position and root it derived itself, and then selects exactly the two reserve leaves for this vault at this generation. A valid proof of another generation, or of another vault by the same owner, selects nothing.

The trader's side

input	source
exact_vault_state_ccb	re-encoded from the composition whose commitment the settle names; the producer re-hashes it and refuses a mismatch rather than assuming
owner_authority_evidence	the composition's own anchor presentation. AnchorPresentationV3 fields 4–9 are AuthorityEvidenceV1 fields 1–6, verified through the same resolver at the same position — the bytes that were checked, not a second format and not a re-fetch
economic_proof_addr	the advertisement's locator, copied and never trusted; the verifier is what makes it mean anything

ComposedVaultState now carries the authority evidence so a consumer takes the authenticated bytes rather than fetching an object that could differ.

Proof

check	result
two devices, production verifier: owner funds legitimately → admitted create → publish; trader with no record of the vault composes from storage, locates the proof, builds the bundle, drives it through the real write-set builder from its own real admitted pre-state, and verify_transition_provenance funds exactly one credit for the curve's output	green
mutation arms, each failing for its own reason: artifact address (proof does not exist), economic position (artifact names another), owner authority (malformed evidence), vault and generation (state no longer hashes to the settle's parent binding)	green
new core controls for gates that had none: proven leaves disagreeing with V_n's reserves; an artifact proving another generation selecting nothing	green
mutation, per rewired gate: artifact verification, V_n agreement, vault-and-generation selection	each reddens only its own named test; restored checksum-verified
core 0x0026 suite	13 / 13
boards	workspace 74 binaries, 3968 passed / 0 failed / 17 ignored (dsm lib 1692/0, dsm_sdk lib 1815/0/7 ignored), exit 0; node 11 binaries, 276/0, exit 0; root make lint exit 0; frontend type-check 0, lint 0, jest 1101 passed; ci/production_safety_checks.sh PASS (pinned 1.98.0, incl. TLA+); scripts/ci_scan.sh PASS — all on tree stamp 8072a355fb88, identical before and after every run

One existing expectation changed, and it is a finding rather than an accommodation: an owner root the proofs do not belong to is now refused earlier and by name, because the artifact declares which root it proves into, where the old message came from discovering the mismatch one leaf at a time. The proposition is unchanged.

Not in this PR

No admission. DlvSettle still advances the device head with no AdmissionPlan and no acceptance fence. The bundle is built and accepted by the verifier; attaching it to a real settle is the next cut.