feat(work-items): run remote comments through continuations

[Neonforge98]

Problem

The durable Work Item dispatcher can hand a remote-root comment to the sender device, but the frontend does not yet claim that offer or execute it through the existing persistent conversation runtime. A transient transport error must also keep the exact durably prepared runner instead of silently rolling to a second Session.

Solution

Mount one app-root Work Item bridge that capability-gates and claims the Rust offer, then delegates all context loading, plane deltas, continuation selection, runner overlay, send/wait/tail publication, and plane signaling to the shared conversation runtime from #944 and #945. Add prepare-before-send and ack-after-accept callbacks to the shared runner, preserve the prepared runner on ambiguous transport failures, and require the deployed retry-safe conversation projection capability. Team Chat remains ordinary human chat and is not moved into the Agent conversation plane.

Potential risks

A cold capability probe can take up to 15 seconds; the parent PR now keeps the offer open for 20 seconds. A crash after durable prepare is safe and retryable, but explicit prepared-runner recovery and native/External CLI runtime selection are intentionally the next stacked PR. Until that layer lands, a missing prepared local Session can remain pending rather than changing runner identity.

Validation

• 33 focused Vitest tests passed
• pnpm exec tsc --noEmit --pretty false
• ESLint passed on all eight touched files
• Normal pre-commit hooks passed

Stacked on #946.

[Neonforge98]

Superseded by the final two-PR design: #844 removes the Work Item-specific remote continuation bridge, and stacked #939 keeps the reusable audience routing plus sender-local provider-native continuation. This dispatcher/store/handoff stack is intentionally retired.