docs(vllm-omni): refresh for v1.5 (0.26.0) — changelog, bidi streaming, s3tokenizer

[Yadan-Wei]

Summary

Refreshes the vLLM-Omni AL2023 DLC docs for the v1.5 / vLLM-Omni 0.26.0 release and documents the new SageMaker bidirectional WebSocket streaming feature. Generated by the vllm-omni-doc-update audit skill (drift + fact-check + coverage-gap), baseline v1.4.0 / omni v0.21.0rc1.

Changes

• changelog/index.md — new v1.5.0 entry (vLLM-Omni 0.26.0 / vLLM v0.26.0): Rust frontend (vllm-rs), SageMaker bidirectional WebSocket streaming, s3tokenizer/CosyVoice3 fix, FlashInfer 0.6.14.
• deployment/sagemaker.md — new Bidirectional Streaming (WebSocket) section: the capability LABEL, the 6 supported WS routes (matched against WS_ALLOWED_PATHS), slashless model_invocation_path, and the inference_ami_version requirement.
• index.md — added s3tokenizer to "What's Included".
• examples/vllm-omni/bidi/deploy_bidi_stream.py — new runnable end-to-end example (deploy → stream TTS over the bidi WebSocket → teardown).

Verification

The bidi example was verified end-to-end against the released prod image (vllm:omni-sagemaker-cuda-v1): deployed a real SageMaker endpoint, streamed 253 KB of audio back over the bidirectional WebSocket, and tore down cleanly. Re-verified under profile-only credentials (no env-var creds) to confirm the credential bootstrap path works for profile/SSO/role users, not just env-var users.

Doc claims were fact-checked against the code:

• 6 WS routes match WS_ALLOWED_PATHS in scripts/docker/vllm/omni_sagemaker_serve.py
• Capability LABEL matches docker/vllm_omni/Dockerfile.amzn2023
• Example request/response shape matches test/vllm-omni/sagemaker/test_sm_omni_bidi_endpoint.py
• Changelog version facts match .github/config/image/vllm-omni/sagemaker-amzn2023.yml

Punch list for a human (NOT addressed here)

• [HIGH — likely a code bug, not a doc fix] EC2 port mismatch. Every example in docs/vllm-omni/deployment/ec2.md uses docker run -p 8080:8080 + curl localhost:8080, but the EC2 entrypoint scripts/docker/vllm/omni_dockerd_entrypoint.sh is exec vllm serve --omni "$@" with no --port, so the container binds vLLM's default 8000. SageMaker sets --port 8080 explicitly; EC2 does not. Either add --port 8080 to the EC2 entrypoint (recommended — makes the docs correct as written) or change every EC2 doc example to -p 8080:8000. Left for an owner to decide since it touches the entrypoint.
• [low] Floating image tag. A couple of doc spots reference vllm:omni-cuda (unversioned) vs the config's vllm:omni-cuda-v1; likely an intentional rolling tag — flagging only.

Test plan

• Verify rendered docs after merge
• Confirm pre-commit run --all-files passes (passed locally)
• Decide on the EC2 port punch-list item