{"content":"docs(1cb348c5): ABS relay fix verification logged — ticket ABS-PERF-2026-1112-01 re-benched against production 2026-11-12 10:12 +08:00 via a fresh e2e identity relay key (primary identity free pool exhausted, resets 11-15). ACCEPTANCE ① SSE passthrough (R1): PASS — a ~130-token answer now arrives in 83/94/87 network chunks with gap P50 0ms / P95 88-179ms (acceptance >40 chunks, P50 <=50ms; before fix: 4-5 chunks at 160-275ms; official baseline same run: 60-69 chunks, P50 10-14ms). Bursty streaming resolved. ACCEPTANCE ② first-token floor (R2): PARTIAL — tiny TTFB improved 1200->685ms median (official 232ms, delta ~+453ms vs target <=+300ms); 4x parallel improved ~1120->558-845ms; but the size-growing component remains: ~4K-token prompt TTFB unchanged at 1280ms median vs official 270ms (+1010ms, FAIL). Residual scales with input size between request accept and upstream first byte — suspect upstream snapshot deployment (deepseek-v4.1-flash-1017) prefill slower than official edge, or per-byte relay body cost. GLM big-prompt delta improved +1805->+595ms vs Zhipu (small-n noisy). Error paths unchanged fast. Remaining ask (supersedes R3): per-phase instrumentation on one 4K-token request or re-point DeepSeek route at official API; acceptance unchanged: success-path TTFB <= official+300ms at both tiny and 4K-token sizes. Addendum in docs/tickets/abs-llm-relay-performance-ticket.md §8.","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}