fix(models): order dual-Ada startup

[nvddr]

Summary

• make shared model-server ordering profile-aware
• start VLM, 8B LLM, STT, then the dedicated 30B LLM on dual Ada
• retain large-model-first startup for single-GPU profiles

Why

The dual-Ada topology benefits from settling shared GPU 0 services before using the dedicated GPU 1 model as the final readiness gate. Single-GPU profiles still need the 30B model to compile against an empty device.

Validation

• uv run --project tests pytest -q tests/test_model_servers.py
• ruff check on changed Python files
• SPDX header check

[nvddr]

Review outcome: ready for review

The branch is rebased on current main, and the earlier topology/test blockers are resolved:

• The dual-Ada default stack now starts VLM and embeddings on GPU 0 before STT and the 30B agent LLM on GPU 1.
• Single-GPU profiles retain their existing 30B-before-smaller-vLLM order.
• The independently selected Omni stack, incompatible-stack cleanup, credential enforcement, embedding configuration, and stack-selection CLI behavior are preserved.
• Tests cover dual-Ada/default, both single-GPU profiles/default, and Omni behavior; documentation now names the current services.

Validation:

• Focused isolated suite: 20 passed (tests/test_model_servers.py)
• Ruff: passed
• Fresh GitHub CI: all checks passed (Python 3.11/3.12, docs build/publish, Ruff, lock, SPDX, DCO, and CodeQL)

No unresolved review threads remain. The PR is ready for maintainer review and merge.

[yanziz-nvidia]

Reviewed by yanziz-reviewer-bot

Summary

Orders dual-Ada startup so VLM and embeddings (GPU 0) settle before STT and the 30B LLM (GPU 1), leaving the largest model as the final readiness gate. Single-GPU profiles (spark, 96G_blackwell) keep the existing 30B-first order. CI green (ruff, pytest 3.11 + 3.12, CodeQL, DCO, docs); DEPENDENCIES.md updated.

Legend: 🚫 Blocker · 💡 Suggestion · 🔍 Nit

	Finding
🚫	None
💡	None
🔍	agent-samples/model-servers/main.py:11-16 — The "Servers started" table still lists stt → agent-llm → vlm → embedding, which now reflects only the single-GPU profiles. A one-line note on the dual-Ada order keeps the docstring authoritative.

Actionables (for bots — copy-paste-ready for AI)

Fix if it makes sense in context — these are agent-generated suggestions, not human-vetted obligations. Skip anything that's wrong, already addressed, or not worth the churn.

• agent-samples/model-servers/main.py:11 — Annotate the "default / --vlm-llm-stack" table to note startup order varies by profile (dual-Ada: vlm → embedding → stt → agent-llm; single-GPU: stt → agent-llm → vlm → embedding), so the header stays accurate after this change.

[yanziz-nvidia]

resolve merge conflicts fist; LGTM otherwise