docs(release): expand Anthropic streaming perf depth-item write-up (v1.87.0rc1)

[yassin-berriai]

Resolves LIT-3334

Stacked on top of yuneng/release-notes-v1-87-0-rc-1 — contributes my depth-item write-up to the v1.87.0rc1 release notes (per @yuneng-jiang's request in #214).

Summary

Expands the Anthropic streaming hot-path perf entry (#28289) — my depth item for this release — from a one-liner into a full write-up:

• The four optimization groups (skip no-op per-chunk work, de-duplicate per-request work, cheaper end-of-stream reconstruction, cheaper hot-path logging).
• Source-of-truth metrics from internal Week-4 real-deployment testing (4-pod m7i.xlarge, no HPA, 256 text_delta chunks/request, validated on both Anthropic direct and Bedrock Invoke): TTFT overhead ~90% lower (p50 2220% → 165%, p95 3057% → 316%, p99 3111% → 328%) and TPM +12% / +6% / +4% (p50 / p95 / p99).
• Notes the wire-output parity testing.

Also leads the matching Key Highlights line with the headline result.

Metrics now come from the internal Week-4 performance doc (real deployment), replacing the local mock-SSE benchmark figures from the PR description.

Not included

• #28794 (management-endpoint SERVER span) — merged 2026-05-25, after this RC was cut, so it is not part of v1.87.0rc1. It belongs in the next release's notes.
• The other PRs I merged this week (#28273, #28362, #28364, #28395) were already documented by Yuneng and need no change.

Test plan

• Docusaurus builds and the note renders at /release_notes

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	May 25, 2026 9:11pm