[FEATURE] Implement Core Agent Fault-Tolerance Endpoints (Start, Checkpoint, Kill, Recover)

Problem

Long-running autonomous AI agents operate blindly across stateless Layer 7 network proxies [1.5]. When an agent pipeline drops middle-turn execution metrics due to network chokes, transient server timeouts, or unhandled errors, the current host memory states are permanently wiped out [1.5, 1.6]. There is no unified infrastructure runtime to orchestrate zero-intrusion checkpoint tracking pools and hot-swap failing processes cleanly without discarding previous token calculations [1.5, 1.6].

Proposed Solution

Introduce a continuous, localized lifecycle execution engine inside internal/gateway/ and internal/storage/ by deploying the following 4 foundational HTTP backend endpoints:

1. POST /api/v1/agent/start: Registers a unique incoming multi-tenant agent workload ID, provisions baseline tracking configurations inside the control block, and targets deployment to active worker instances.
2. POST /api/v1/checkpoint/create: Performs synchronous snapshots of state arrays and bit frames, committing memory context snapshots dynamically onto our atomic, in-memory circular storage queues (CacheManager).
3. POST /api/v1/process/kill: Artificially fires a process interruption signature or network stack exception to simulate an abrupt serverless container crash event or upstream model endpoint failure.
4. POST /api/v1/state/recover: Intercepts the crash signal, handles error masks, pulls the last available checkpoint registry bits, and gracefully resumes execution on healthy standby targets in sub-2ms boundaries [1.5, 1.6].

Alternatives

Managing state retention inside high-overhead userspace database grids, which was completely rejected due to heavy string serialization latencies and processing thread allocation blocks.

Use Case

Secures direct end-to-end resilience validations across long-running multi-turn automated workforces, letting enterprise developers micro-benchmark failover recovery metrics reliably under chaotic system scenarios [1.5, 1.6].

Additional Notes

Implement logic routines cleanly in Go utilizando native synchronization primitives (sync.RWMutex) to guarantee safe concurrency bounds without garbage collection overhead [1.5].

Start by reading the existing code under internal/gateway/ and internal/storage/, including the CacheManager mentioned in the issue. Define how the four lifecycle endpoints share workload and checkpoint state, then verify that concurrent start, checkpoint, kill, and recovery flows preserve the last checkpoint and meet the stated recovery target.