[Cross-vendor] Replace MI300X measured template with real AMD artifact

[Darkroom4364]

Context

Cross-vendor transfer has a prediction scaffold and a measured-results template, but the MI300X measured artifact is not real evidence yet. The current docs/results/cross-vendor-measured-mi300x-template.json has placeholder measurement fields by design.

Scope

• Run the cross-vendor protocol on real MI300X or MI250 hardware, or explicitly document why measured AMD data is deferred.
• Produce docs/results/cross-vendor-measured-mi300x.json with real metrics.
• Run scripts/cross_vendor_transfer_eval.py against the prediction and measured artifacts.
• Write the JSON/MD evaluation outputs and link them from docs/results/README.md.

Acceptance criteria

• No paper-facing wording treats the scaffold/template as measured AMD evidence.
• The produced evaluation reports Spearman correlation, top-k overlap, and latency regret.
• If hardware access is unavailable, docs clearly mark the cross-vendor lane as scaffold-only.

Related: #82.