[WIP] feat(router): add engine-pressure-aware scheduling

[yangbodong22011]

Motivation

Define Engine pressure-aware scheduling independently so it can be reviewed while the LoadMonitor dependency remains in #32510.

This is a stacked WIP that depends on #32510. No #32510 implementation is included in this PR.

Modifications

• derive normalized pressure from request saturation, queued uncached tokens, prefill/decode throughput, and hottest-rank KV usage
• route load_based candidates using inverse-pressure weights
• assign zero routing weight to fully pressured engines
• preserve least-active routing when Engine metrics are unavailable
• add deterministic scorer and weighted-routing tests

Accuracy Tests

Not applicable. This change affects Router backend selection, not model execution or output accuracy.

Speed Tests and Profiling

Pressure scoring is O(1) per candidate and weighted selection is O(n). No request body buffering or model-path synchronization is added.

Tests

Validated after applying this patch to the exact #32510 head (5528529ab3):

• cargo fmt --all -- --check
• 4 pressure scorer tests passed
• 4 load-based scheduling tests passed
• cargo clippy --all-targets -- -D warnings

The standalone WIP intentionally does not compile until #32510 provides AggregateLoad and PolicyCandidate.

Checklist

• Format code
• Add unit tests
• Follow SGLang code style guidance
• Update documentation if the final user-facing configuration changes
• Run full benchmarks after the dependency is integrated

---

CI States

Latest PR Test (Base): ✅ Run #30912391345
Latest PR Test (Extra): ❌ Run #30912390880

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.