v0.6.27

• chore(release): refresh locks and generated sources
• chore(main): release 0.6.27
• fix(server): constrain SGLang compatibility paths
• fix(server): preserve LightOn SGLang compatibility
• fix(server): keep Qwen3-VL reranker batches working on transformers 5.3
• fix(generation): pair long-context qwen image bounds
• fix(generation): preserve qwen vision bounds
• fix(generation): validate high-context thinking profiles
• test(generation): align grammar backend expectations
• fix(generation): align direct backend output limits
• fix(generation): use compatible Qwen 27B grammar backend
• fix(generation): use compatible Qwen 35B grammar backend
• feat(generation): add hardware-specific context profiles
• fix(generation): preserve Gemma reasoning boundaries
• fix(generation): fit Gemma 256k KV cache on H100
• fix(sdk): retry pre-execution SSE capacity errors
• fix(generation): reserve Gemma long-context workspace
• feat(generation): add 256k profile candidates
• feat(generation): add strict Responses API parity
• fix(generation): hide split Gemma thinking boundary
• fix(generation): reject overflowing runtime numerics
• fix(generation): handle family-specific reasoning boundaries
• fix(generation): unblock cold model startup
• fix(server): align Gemma xgrammar dependency
• feat(server): complete generative model serving
• fix(server): enforce Qwen fast image bounds
• test(server): align vision profile contracts
• fix(server): bound Qwen visual context
• fix(server): replace stale root playground with redirect to Swagger docs
• fix(server): pin docling to its artifact revision and stop OCR gating load
• fix(colbert): honor transformers-5 rope_parameters when serving on the 4.x bundle
• fix(cloud): close connector review gaps
• fix(cloud): close connector review findings
• fix(clients): send connector repair idempotency header
• fix(cloud): address connector runtime review findings
• test(cloud): prove public postgres connector lifecycle
• feat(clients): expose recovery-bound connector repair
• feat(cloud): add direct connector inference
• feat(gateway): expose immutable execution evidence
• feat(clients): add connector plan controls
• feat(cloud): add connector plan schema policy
• feat(cloud): add connector plan guardrails
• fix(models): complete mLateOn serving recipe
• fix(server): preprocess native extract images
• feat(models): onboard lightonai/mLateOn (scaffolded)
• fix(sidecar): harden local generation cancellation
• test(server): shorten local ingest socket paths
• fix(sidecar): reserve cancel request bytes
• feat(server): add bounded local-ingest generation client
• fix(sidecar): harden local generation streaming
• feat(sidecar): stream generation over local ingest
• feat(models): onboard ibm-granite/granite-embedding-97m-multilingual-r2 (scaffolded)