docs(deploy): self-hosted evaluation guide + accurate GPU requirements

[ghchinoy]

Summary

Answers three recurring customer questions (hosting TL;DR, required GPU, how to evaluate) with one page and corrects an overstated GPU claim.

New: docs/deploy/evaluate.md — "Evaluate dgem on Your Own GPU"

Ordered steps, each with a pass criterion:

0. Hardware: RTX PRO 6000 (recommended, tested); L4 works (~3× slower; DISABLE_MM=1 on 32 GiB host RAM, 64 GB with vision on); other GPUs untested.
1. Readiness: docker run + /health until vllm_ready and warmed; how to mount pre-downloaded weights.
2. First decision: dgem decide --stats -v samples=1; expect ~55–65 ms Server Denoise on RTX PRO 6000.
3. Baseline: bench-jev and bench-calibration at -w 4; pass ranges 183–192/231 and 43–45/50 (16 runs each, benchmarks/runs/20260927-image-parity).
4. Concurrency (optional): same suites at production concurrency, 0 errors.
5. Custom evaluation: policy → labelled dev/test split (100+ per question) → local dgem serve -u <GPU_HOST> for Studio Batch Eval or the Python runner → scripts/policy_calibration.py → what to record.

Mirrored to docs-site/, added to the "Build, Deploy & Operate" sidebar, linked from docs/index.md, index.mdx, deploy/index.md, deploy/laptop.md and deploy/public-images.md.

Corrections

• deploy/public-images.md: section heading claimed "Any GPU Host (NVIDIA Blackwell, Ada, Hopper, Ampere)"; we have no receipts for the 4-bit image on Hopper/Ampere. Replaced with a tested/works/untested table. Clarified DISABLE_MM.
• reference/vertex-vs-cloud-run.md: A100/H100 listed as Vertex machine types now notes the public images are tested only on RTX PRO 6000 and L4.

Verified

• dgem serve -u http://<non-loopback-host>:8080/v1 alone routes to that host (available_backends: ["cloudrun"]); a loopback URL routes as local. Both work for /api/decide.
• Gateway does not forward -k tokens to loopback URLs (cmd/serve.go), hence the guide's API_KEY note. Possibly worth a follow-up fix.
• make docs-sync-check (44/44), scripts/docs_link_check.py (0 broken), make check-public pass; docs-site astro build succeeds (45 pages) and in-page anchors used by the guide resolve.

Not verified

• The guide's steps were not executed end to end on a fresh self-hosted GPU in this change; the numbers come from existing receipts.