Blob model

This is the planned guest blob capability model. The completed Phase 2 node stores verified packages and maintains an optional bounded raw artifact cache; neither grants guest blob access. Its invocation path uses bounded inline payloads and rejects unavailable blob imports. Phase 3 plans local and S3 blob providers with bounded writes, reads and leases. Cluster replication remains later work. See the capability surface and roadmap.

Large payloads should not be repeatedly serialized through the router, runtime, and component call graph. The intended model represents them as immutable content-addressed blob references.

Write lifecycle

begin staged write
  → bounded ranged writes
  → digest and size verification
  → seal immutable blob
  → issue activation-scoped lease

A failed or expired write session is reclaimable and cannot become visible as a valid blob reference.

Read lifecycle

The capability broker grants a lease limited by tenant, activation, operation, byte budget, range, and expiration. Same-node implementations may use memory mapping or shared immutable regions; remote calls may transfer by digest. These choices cannot expose raw unrestricted host pointers to guest code.

Ownership and retention

Blob identity is content digest plus tenant policy. Retention, pinning, replication, and garbage collection are platform concerns independent of capsule execution lifetime.