[knapcio]

Summary

• backport the runtime part of SGLang #44001 to the pinned f80c91a4b tree with full upstream credit
• keep the new --prefill-decode-interval-min-bs option at zero by default, preserving current scheduling
• add zero-fuzz staging, semantic verification, build-time tests, CI coverage, and opt-in operator docs

Why

With --prefill-decode-interval 2, a small running batch can delay queued requests before the interval is useful. The new optional threshold admits prefills eagerly below a selected live batch size and restores the interval at/above it. This is a multi-user TTFT candidate, not a C1 decode-TPS claim.

Verification

• fresh f80c91a4b GitHub tarball: #41303, #42901 and #44001 applied with patch --fuzz=0
• all three semantic verifiers passed on the staged tree
• offline regression gate: 54 tests passed
• Python compile, shell syntax and git diff --check passed

No fleet, container, NUC, network or live serving state was touched. GB10 A/B remains required before enabling min-bs in a deployment.

Credits: Shangming Cai / sgl-project/sglang#44001.