Background

PR #451 adds DeepSeek-V2-Lite HTTP trace evidence for issue #280, and issue #452 follows up on mixed and long-prompt latency. Those cover completion, traceability, and a narrow performance path, but they do not yet prove that the HTTP serving path stays healthy under common production-style failures.

The current scheduler work already has unit coverage for EOS, disconnect, cancelled sinks, unsupported requests, and batch decode errors. The missing gate is an HTTP-level reliability run that proves those behaviors hold under real request pressure.

Current state

• Mixed-request admission and per-request KV ownership are covered by dsv2lite: add mixed-request serving admission and KV ownership #281.
• HTTP trace and decode subgroup evidence are covered by dsv2lite: diagnose why HTTP serving throughput does not scale with concurrency #280 / PR DeepSeek-V2-Lite: HTTP concurrency evidence, subgroup decode batching, and long-prompt fix #451.
• Mixed and long-prompt TTFT/ITL optimization is tracked by dsv2lite: reduce mixed and long-prompt HTTP TTFT with prefill scheduling #452.
• We still lack a retained HTTP reliability gate for client cancellation, disconnect, overload, and bad requests.

Desired outcome

Add a narrow DeepSeek-V2-Lite HTTP reliability gate that shows the server can reject, finish, cancel, or recover requests without poisoning unrelated in-flight requests or leaking scheduler state.

Acceptance boundary

• Run HTTP reliability checks against DeepSeek-V2-Lite EP2 with the intended server path.
• Cover at least:
  • client cancellation or disconnect during streaming;
  • unsupported or invalid request parameters;
  • overload / active-cap pressure with queued requests;
  • mixed short/long prompts while the above failures occur.
• Confirm normal requests continue to finish when neighboring requests fail, cancel, or disconnect.
• Record completed, failed, rejected, cancelled/disconnected, and timeout counts.
• Record output hash stability for successful requests when applicable.
• Record trace coverage and request retirement reason for terminal requests.
• Confirm active and pending request state returns to a healthy baseline after the run.
• Preserve HF / host-staged / NCCL exactness for the committed DeepSeek-V2-Lite case set when runtime behavior is changed.

Boundary

This is a reliability gate, not a performance claim. It does not claim vLLM parity, production EP readiness, sparse dispatch readiness, multi-node EP support, or a final runbook.

Related roadmap: #205
Related diagnosis: #280
Related follow-up: #452
Related PR: #451