Vision enablement evidence

Functional acceptance on two RTX PRO 6000 Blackwell Max-Q 96 GB GPUs, native Linux.

• Managed recipe reconstruction: synthetic model-cache mount; operator GPU ownership and endpoints remain private.
• Direct acceptance, image corpus, long-context checks, routed acceptance.
• Real Pi image response, Mini client sync, repeat sync.
• Decision, final state, friction, publication copy.

Raw schemas and results are preserved. Absolute operator evidence/product paths were replaced with EVIDENCE_ROOT/PRODUCT_ROOT; the private DNS name was replaced with example.ts.net. The runtime build ref is inherited from the pinned baseline metadata; startup reports vLLM 0.1.dev20111+g7f1e92bec.d20260827. No new source-to-image attestation was performed.

Additional negative evidence: 259,922-token retrieval refusal; the earlier 231,045-token retrieval passed.

User-requested retry: the same 259,922-token fixture passed after the user reported an overlapping prompt. Both attempts are retained; concurrency causation is not established.

Later eight-image follow-up: four ordinary-size requests and one eight-image 2520×2520 request passed, with a matching Pi output transcript (uncorrelated request provenance). The large request used 63,456 input tokens and completed in 28.17 seconds. Limit now eight/request; no simultaneous eight-image soak. Native evidence, large request, summary.