feat: add GPU runtime upgrade skill

[jiaran-king]

Summary

• add an upgrade-gpu-version skill for upgrades across pinned vLLM revisions and exact CUDA runtime/toolchain environments
• gate implementation on immutable current/target identity, upstream contract analysis, and major-architecture-change detection
• separate planning, implementation, and final-validation modes while reusing the existing run-e2e skill for GPU execution
• add a reusable upgrade workbook, historical v0.19.1-to-v0.26.0 lessons, and a documentation refresh gate/checklist

Why

GPU upgrades span compatibility patches, models, workers, model runners, connectors, distributed topology, CUDA Graph, DBO, profiling, native ABI, packaging, and the runtime software stack. This workflow makes upgrade evidence reproducible, stops unsafe local adaptation when upstream makes a major architecture change, and keeps public support claims aligned with exact validation evidence.

Validation

• quick_validate.py .agents/skills/upgrade-gpu-version
• git diff --check
• repository pre-commit YAML, end-of-file, trailing-whitespace, and merge-conflict checks
• read-only planning-audit dry run for [RFC]: Upgrade the GPU backend to vLLM 0.26.0 #167 against exact vLLM v0.19.1 and v0.26.0 sources

Closes #224

[jiaran-king]

Dry-run validation

The upgrade-gpu-version skill completed a read-only planning-audit dry run for the vLLM 0.19.1 → 0.26.0 upgrade described in #167.

Frozen identity

• AFD baseline: release/v0.19.1rc1 at 2f4f4627d6606dc8b6b11b406d3bc5b31e363090
• Current vLLM: v0.19.1 at b1388b1fbf5aaef47937fabe98931211684666a6
• Target vLLM: v0.26.0 at 568afb3a13806beb53bb2e6bd518269357b237c0

Phase results

• FREEZE_IDENTITY: PASS
• INVENTORY_AFD_CONTRACTS: PASS
• DIFF_VLLM: PASS
• ARCHITECTURE_GATE: PASS
• PRODUCE_STAGED_PLAN: PASS

The dry run produced:

• a GPU/shared AFD surface inventory;
• compatibility-patch and private-seam dispositions;
• an upstream-to-AFD impact matrix;
• an architecture-gate decision;
• a staged implementation and validation plan.

The architecture gate passed because vLLM 0.26.0 retains the V1 worker/model-runner lifecycle required by the existing AFD GPU design. ModelRunner V2 was kept explicitly outside this upgrade scope. The plan also identified opportunities to delegate DeepSeek model lifecycle and weight loading to target-native implementations and to reconstruct compatibility patches from the target source rather than incrementally modifying old copies.

The AFD and vLLM worktrees remained unchanged throughout the dry run. This result validates the skill’s complete planning workflow; it is not an implementation or release-readiness result.