Part of epic #3602.

Scope

Implement a small round-robin pool of gonka node endpoints with a fail-skip cooldown. No tower-balance, no health-checking daemon — those are explicit non-goals.

Files to create

• crates/zeph-llm/src/gonka/endpoints.rs:

  pub struct GonkaEndpoint { pub base_url: String, pub address: String }
  pub struct EndpointPool { nodes: Vec<GonkaEndpoint>, cursor: AtomicUsize, failed_until: Vec<AtomicU64> }
  impl EndpointPool {
      pub fn new(nodes: Vec<GonkaEndpoint>) -> Result<Self, LlmError>;  // empty = error
      pub fn next(&self) -> &GonkaEndpoint;                              // round-robin
      pub fn mark_failed(&self, idx: usize, cooldown: Duration);
      pub fn len(&self) -> usize;
  }

Behaviour

• next() returns the next non-failed endpoint via AtomicUsize cursor.
• mark_failed(idx, cooldown) stores now + cooldown in failed_until[idx]; next() skips entries whose cooldown has not expired.
• Cooldown default 30 s.
• If every endpoint is in cooldown, next() falls back to the least-recently-failed one.

Acceptance

• Inline unit tests in endpoints.rs cover:
  • Round-robin order over 3 nodes.
  • Failed node skipped during cooldown, restored after.
  • All-failed fallback returns a valid endpoint.
  • Empty constructor errors.
• cargo nextest run -p zeph-llm -E 'test(gonka_endpoint)' green.
• No mutex usage (per await discipline rules); only atomics.

Depends on

None.

Size

S (~2h)