Note

Low Risk
Display-label only; omitting model on getHardwareConfig keeps prior MTP behavior, with tests covering cache isolation.

Overview
MiniMax-M3 runs stored with spec_method / hwKey token mtp now show EAGLE (e.g. suffix (vLLM, M3 EAGLE)) instead of generic MTP. Other models are unchanged.

Adds resolveFrameworkPartLabel and MODEL_SPEC_METHOD_LABELS in @semianalysisai/inferencex-constants, and extends getHardwareConfig(hwKey, model?) so suffix parts respect per-model overrides. Model-scoped cache keys avoid leaking M3 labels into model-less lookups.

Call sites pass the display model (or row model) into getHardwareConfig across inference charts, tables, CSV export, eval labels, submissions, changelog config keys, calculator throughput, and AI chart bars. SubmissionsTable and eval/changelog builders use the same resolver for spec-method cells and headers.

^{Reviewed by Cursor Bugbot for commit 1c96c70. Bugbot is set up for automated code reviews on this repo. Configure here.}