Add idea: durable execution as a CF-native primitive

[rkoster]

Argues that CF should add a durable execution primitive, grounded in what the agent frameworks surveyed in this repo actually need \u2014 and notably, what none of them ship.

The gap: cf run-task runs once (no checkpoint, resume, retry, or schedule); app instances run forever with no durable per-instance state. Durable execution lives in between, and CF has nothing there. Every framework surveyed delegates durability to a substrate: ADK exposes replay-safe primitives but no engine, OpenAI Agents SDK relies on external Temporal/Dapr, Microsoft AF binds to Azure Durable Task, LangGraph/CrewAI/LlamaIndex all expose pluggable checkpointers.

Key positions:

• CF owns the lifecycle; a bound service owns the state. Apps already checkpoint correctly today (LangGraph PostgresSaver against a bound Postgres). CAPI never becomes a database \u2014 it stores a pointer.
• The missing verb is "I've checkpointed \u2014 release my container, wake me on event E or at time T." A suspended execution should hold no container; today an agent awaiting human approval for three days occupies a full app instance.
• Tasks are the right execution primitive but the wrong identity primitive. Diego Tasks are already short-lived and release their container on exit \u2014 the right shape for a slice between suspensions. But task GUIDs are minted per run and tasks can't be addressed, so identity moves to a durable CAPI resource above them (mirroring Temporal's workflow-execution vs. workflow-task split).
• One API per cell, not per container. The Loggregator agent is the precedent; a durable-execution API is a thin stateless control call, unlike a Dapr sidecar which must be per-container. Callers authenticate via mTLS with the Diego instance identity cert, reusing RFC-0055's identity extraction.

Explicit non-goal, with evidence: no HTTP-addressable resume. An evidence table across Temporal, LangGraph, LlamaIndex, CrewAI, OpenAI Agents SDK, Letta, and Cloudflare shows durable executions are near-universally addressed by stable ID through an always-on service, never as a network endpoint \u2014 Cloudflare being the sole exception, because a Durable Object is simultaneously compute, state, and addressing unit. This avoids needing a Knative-activator-style request-holding component.

Deliberately argued on framework evidence alone rather than as a counterproposal to the sibling Dapr ideas (#32, #33).