fix(stargate): bound proxy retries and relay retention

[barrygreengus]

Why

A response-header timeout could replay a POST after the backend had accepted it. Long-lived QUIC relays retained completed tasks, and arbitrary unknown model/routing-key headers created permanent metric series.

What changed

• Track whether each attempt handed its request body to the tunnel, and report ambiguous delivery instead of retrying it. Preserve reconnects before submission and retries after explicit pylon rejection.
• Reap completed tasks before accepting more streams in both normal and draining relays.
• Aggregate unknown-target 404 requests under one pre-initialized metric series.
• Update the request sequence diagram and add an end-to-end duplicate-execution regression.

Customer Release Notes

Stargate avoids duplicate inference requests after uncertain transport failures and bounds memory retained by relay completions and unknown-target metrics.

Plan Summary

Not applicable.

Usage

A transport failure after body submission returns the upstream transport error to the caller. Explicit retryable pylon rejection behavior is preserved.

Testing

The rebased stack tip passed all 1,611 Cargo workspace tests, including 143 Stargate integration tests, with 3 existing ignored performance tests. Workspace Clippy passed with warnings denied. The POST regression confirms one backend execution and a 502 response after a header timeout. Relay lifecycle and bounded unknown-target metric regressions pass. No live-cluster or production traffic soak was run.

Notes

Layer 2 of the Stargate maintenance stack. Unknown-target 404 metrics use empty routing_key, model, and inference_server_id labels; raw rejected identities remain in logs.

Issues

Relates to #1817

References

HTTP retry semantics

Related Pull Requests

Depends on #1818. Next layer: #1822.

Dependencies

No added or upgraded dependencies. No license or NOTICE changes.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

🛡️ CodeQL Analysis

🚨 Found 5 issue(s)

Severity Breakdown:

• 🔴 Errors: 0
• 🟡 Warnings: 0
• 🔵 Notes: 0

📋 Top Issues

🔗 View full details in Security tab

_{🕐 Last updated: 2026-09-11 21:22:52 UTC | Commit: 2d1ee42}