The Missing Knowledge Layer for Agentic EDA: Versioned Tool Documentation, LLM-Mined Ontologies, and Live Project Memory for Autonomous PPA Closure

Autonomous LLM agents are increasingly completing chip design flows end-to-end, especially on the open-source OpenROAD stack. Their most common failure is procedural: the model writes a flow variable or value that looks plausible but does not exist, or is illegal, in the installed tool version. Pasting documentation into the prompt as static ``skill'' files inflates the context and is unreliable in published evaluations, while hand-built tool gateways work but must be re-engineered per tool. We supply the missing layer: the installed tool's own documentation and configuration schema, at its exact pinned version, compiled into a typed knowledge graph the agent queries while working, validates every edit against, and grows with each verified fix, so subsequent runs retrieve past solutions instead of re-deriving them. On power, performance, and area (PPA) recovery tasks on ASAP7, a plain 31B-model agent matches published frontier results; adding this layer turns two of its five power-task failures into full successes and lifts mean recovery on these tasks from 79\% to 94\% of the degraded gap, cuts actions-to-success by 38\% on tasks both solve, and drives invalid edits to zero.