[matthewgoodman13]

Description

Amazon ECS GPU sharing with NVIDIA Multi-Process Service (MPS) launches on 2026-10-22. Without guidance, an agent asked to pack small models onto fewer GPUs tells the user ECS can't share a GPU, recommends the older manual method (remove GPU resource requirements and make the NVIDIA runtime the default through instance user data), or suggests a multi-model inference server. This PR adds a short reference so the agent recommends managed GPU sharing when it fits and reads the official docs before advising.

Changes, applied identically to skills/core-skills/aws-containers and plugins/aws-core/skills/aws-containers:

• New references/ecs-gpu-sharing.md. It covers:
  • What GPU sharing does: a sharingStrategy with an nvidia_mps object in the container's GPU resource requirement, with a GPU memory limit in MiB and an optional maxcomputePercent. ECS places tasks by GPU memory only.
  • When to use a whole GPU instead (more than one CUDA process per container, or per-task GPU metrics) and where sharing does not work (Fargate, ECS Express Mode, managed daemon tasks, G6f fractional GPU instances).
  • Security considerations: the shared GPU failure domain, memory as a hard ceiling, and avoiding the manual method that lets every container on the instance use the GPUs.
  • A requirement to read and cite Share GPUs on Amazon ECS. It copies no validation rules, examples, or instance lists.
• references/ecs.md: one routing line to the new reference.

SKILL.md frontmatter is unchanged, so skill selection is unaffected. This PR touches the same ecs.md routing list as #476 (ECS Gateway); whichever merges second needs a one-line rebase.

Validation (internal, pre-launch, docs not yet public):

• End-to-end knowledge case, 5 models x 3 runs (Sonnet 5, Opus 4.8, GPT-5.6, Opus 5.5, Sonnet 5.5): 13/15 (87%) with the skill vs 0/15 (0%) without. Both misses lost only the documentation citation.
• Omega CR eval (Sonnet 5, full aws-containers suite): 50/54 (93%) with the skill vs 11/54 (20%) without.
• Skill-selection eval, 972 queries, 6 models: 98.6% to 99.4%, with 0 aws-containers misses beyond one unrelated GPT-5.4 miss that did not repeat.

Type of Change

• New plugin
• New skill
• Bug fix
• Documentation update
• CI/CD change
• Other

Checklist

• I have read the CONTRIBUTING guide
• My changes pass python3 tools/validate.py
• I have updated relevant documentation

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.