feat(openengine): add shared engine-neutral sidecar

[connorcarpenter15]

Summary

• add one engine-neutral dynamo-openengine-sidecar binary for TRT-LLM, vLLM, and SGLang
• negotiate compatible OpenEngine revision ranges and discover model aliases, tokenizer, role, handoff profile, multimodal, LoRA, load, and KV-event capabilities
• forward opaque context-first handoffs, coordinate typed SGLang bootstrap, bridge ZMQ or gRPC KV events, and preserve media-aware routing keys
• close Generate streams on cancellation, issue idempotent Abort, and drain admitted streams to terminal completion
• add aggregate/disaggregated launch examples, DGD manifests, and parameterized fake-server coverage for all three engines

CPU validation

At Dynamo head decfa3781be5134efef11f444f22bcff21fbfc92:

• cargo fmt --all -- --check
• cargo check -p dynamo-openengine-sidecar --locked
• 40 shared-sidecar tests
• 11 focused backend error/stream tests
• shell syntax and DGD/YAML parsing checks

Exact shared artifact

Every final computelab run used the identical binary:

• sidecar SHA-256: d7b1205a948fbc1636322d070816248f8c0d5a3f9554792deca6700611e7deb5
• manifest SHA-256: 7ffd6b25b8e04c63e3115a00feada451c52047507692748eaaeba1b4391a77ae
• ELF build ID: fdba81f7e1aa93c66ab302290f5263115ac866be
• OpenEngine/schema release: a66ff6f73a65e262a7c3edd5ea6fd0d8701d402f
• TensorRT-LLM: 58132ce1f654bfede8e530f78e31979596c5c34e
• vLLM: 5bab3696d889a28b632b8a54d97ff32d0c982697
• SGLang: f5c8f597dfb9be027276a7bacdac36fd7f988de1

GPU validation

Computelab used the same directly connected H100 NVL pair for the serialized final campaigns.

• TRT-LLM: 8/8 phases pass—aggregate lifecycle, event-backed and attention-DP routing, text P/D, image/video aggregate+routing, image/video P/D, audio aggregate, and audio-P/D rejection
• vLLM: aggregate lifecycle, event-backed routing, DP=2 routing, image aggregate/P-D/routing, and unsupported-capability rejection pass; a corrected focused text-P/D run proves a correlated 192-block NIXL handoff, nonzero admitted Drain, typed new-work rejection, completion of 16 admitted streams, failover, and clean exit
• SGLang: aggregate lifecycle/routing, DP=2 routing, text P/D, image/video aggregate/P-D/routing, audio aggregate, and unsupported-role rejection pass; the focused same-budget text-P/D rerun exits every process within the 20-second grace

The original vLLM drain timing false negative and SGLang 20.222-second teardown miss remain preserved. Their corrected focused validations reused the exact artifact and product commits.

Fork CI limitations

Local CPU and computelab product gates pass. Remaining red fork workflows are infrastructure/policy-only:

• codeowners evaluates the non-ancestor fork merge-base and flags unrelated upstream squeeze_evolve files
• Copyright Checks cannot pull ghcr.io/connorcarpenter15/dynamo/helm-tester:0.1.1
• reusable Rust jobs are created with no runner and execute zero steps

Docs, DCO, CodeQL, pre-commit/title/label gates that can run on the fork pass.

Related PRs

• OpenEngine #20
• TensorRT-LLM #1
• vLLM #15
• SGLang #5