Tracks #67, #68, #69, and #70. Independent reproduction: ruvnet/metaharness#301.

Finding

Beyond Agent Harnesses: Cross-Substrate Authority for Multi-Agent Systems (arXiv:2609.08472, submitted 2026-09-08) isolates a composition failure where authorization state lives outside planner-visible workspace and memory. The originating study reports 0/32 final semantic success for authority-blind candidate evidence versus 32/32 when the missing authority fact is visible. Typed packaging gives no observed planning advantage over equivalent raw evidence. Planning remains unreliable even with authority context, while a deterministic execution guard blocks all six replayed unsafe intents and permits all 12 valid authorized publish intents.

These are external originating-team measurements, not RuV benchmark claims.

Change

Adds ADR 163 and a dependency-free, allocation-free, no_std strict continuity verifier to rvm-cap.

The verifier binds eight decision-relevant authority fields across a component transition:

1. authenticated principal
2. task scope
3. provenance root
4. capability scope
5. active policy state
6. canonical action
7. effect boundary
8. capability epoch

Every field must remain identical. Planner-visible input and output context may evolve but are bound into the receipt. Any authoritative change fails closed.

The returned receipt explicitly carries ReceiptAuthority::None. It does not consult the capability table, mint rights, refresh an epoch, approve an effect, or provide a caller-controlled release bit. Existing RVM capabilities and permits remain authoritative.

Tests

Adds deterministic coverage for a valid unchanged transition plus principal substitution, task rebound, provenance discontinuity, capability scope change, policy change, action reinterpretation, effect-boundary change, stale epoch, and advanced epoch.

Benchmark

MetaHarness #301 freezes at least 32 fault classes across MCP to RVM, Ruflo delegation to RVM, Core Memory retrieval to RVM, and one lifecycle boundary. Promotion requires zero harmful effects, benign completion within 2 absolute points, all changed authority context failing closed or escalating, p95 local verification below 1 ms, and zero capability expansion.

Security and rollback

No new dependency, allocation, unsafe code, network, filesystem, credential, or persistence surface. Durable or remote receipts must be sealed by the existing RVM or RVF witness path. Rollback removes the additive module, export, and integration call sites; no state migration is required.

Keep draft until hosted CI, clippy, no_std, security/audit gates, and MetaHarness #301 complete. No autonomous merge or deployment.