Assistant apps: expose Inbox operations through the scoped app SDK

Related: #1645, #1646, #1570.

The personal assistant interprets intent across conversations, Inbox and work, while a specialist agent may have a narrower objective and tool scope. Both can use the same harness. Single-user versus shared use is a separate dimension, not an inherent limitation of agents.

Audit existing Inbox HTTP handlers and the apps sandbox bridge before adding endpoints. HTTP access alone is not a usable app integration: provide a small supported SDK over the same authorized backend operations, analogous to the existing mu.agent.stream surface.

• Inventory existing operations and missing bridge/SDK coverage; reuse current thread/work semantics rather than build a second Inbox.
• First slice: list owned threads with pagination, read one and continue/reply where supported; then assignment and work status using existing authorization and delivery.
• Specify stable result shapes, errors, cancellation and pagination; keep credentials in the parent/runtime, not app source.
• Apply the same account and resource authorization at the server on every route; the iframe bridge is not the only security boundary.
• Exercise one Assistant app flow: find a conversation, read it, continue or assign explicit work and show the actual result/status.
• Preserve the chat-first app experience; do not automatically embed every service card or reproduce Home navigation.
• Keep private defaults and make source/audience visible. Group access follows the shared membership model rather than passing arbitrary owner IDs.

Queue after the first reliable app-maintenance loop in #1646. This is incremental product discovery; no wholesale replacement of built-in UI.

Start by auditing the existing Inbox HTTP handlers, apps sandbox bridge, and the mu.agent.stream surface. Inventory the supported operations and missing coverage before defining the SDK's result shapes, errors, cancellation, pagination, and authorization behavior. Done means an Assistant app can find and read a conversation, continue or assign explicit work, and show the actual result or status.