fix: runtime RDMA providers and native hybrid-state support (v2.27.0)

[ketor]

Changes

• Include ibverbs-providers in the shipped runtime image and add an RXE-probe-gated real RDMA runtime-container smoke to CI.
• Refresh the version-scoped native SGLang hybrid Mamba/DSA patch: retain indexer sidecars and the allocator's actual scaled-MLA geometry.
• Ship a targeted native vLLM GLM Flash V2-runner fix: bypass generic position-indexed slot mapping for KpoolTailSpec; the existing circular metadata builder still supplies real tail mappings. Retain the existing allocation policy.
• Add an explicit, bounded real-CUDA L1/L2 offload/poison/restore regression and document scope and patch applicability.
• Synchronize VERSION, all connector wheel versions, and exact common-wheel pins to 2.27.0. The runtime packaging fix is server-side; native C ABI, wire formats, and connector runtime code are unchanged.

Verification

• Repaired runtime image: real RDMA PUT/EXIST/GET smoke, including larger objects with the actual RDMA object-bound configuration.
• Native GLM Flash: reproduced an unconnected engine failure in the slot-mapping kernel; the minimal patch passes the same short/short/long sequence and native baseline phases without disabling prefix caching.
• Real-CUDA regression: the original hybrid assembler fails actual restored bytes; the patched assembler passes both raw-FP8 and scaled-MLA layouts, including indexer and recurrent/convolution state.
• Framework/model runs cover cold writes, local-cache baselines, external-only cross-process cache restoration, byte comparisons, and uninstrumented repeated performance measurements.
• All three local wheels build with 2.27.0 metadata and exact matching common-wheel dependency pins.
• Independent static reviews found no actionable defects; CI remains a required merge gate.

Acceptance limits

These are pinned native-image integrations, not a claim that generic stable images support every model. Full generated content/reasoning equivalence remains unresolved, including repeated no-L3 baselines; it is not waived or reported as passed. Cache-byte/targeted-answer evidence is reported separately. Acceleration is workload-dependent, with weak/negative cases and first-use outliers retained. The bounded GPU regression does not replace distributed/L3/model/MTP or speculative-state acceptance.