loujc/team20_mlcad_draft_agentflow/mlcad-erc-buffering

MLCAD26: ERC-Aware Buffering & Cloning

Use this skill only when current-run evidence shows ERC or high-fanout bottlenecks.

Net Selection

Rank nets by:

• slew/cap/fanout overflow amount;
• fanout count and load distribution;
• driver strength and load class;
• bounding-box/HPWL proxy;
• timing criticality and endpoint cluster;
• boundary risk.

Hard-reject nets touching top ports, clocks, reset/set controls, DFF/latch/ICG outputs, SRAM/macro outputs, tie/tap/physical-only cells, or protected endpoints.

Repair Options

Prefer the smallest safe repair:

• size the internal driver if it fixes slew/cap without power blow-up;
• insert a buffer on a safe internal segment when fanout/cap requires it;
• clone a driver only when net partitioning is clear and legal;
• remove redundant buffers when timing/ERC remain clean and dynamic power improves;
• run local DPL/check_placement after accepted topology edits.

Iteration

1. Start with a tiny probe: 1-10 nets or a bounded native repair command with protected sets.
2. Evaluate ERC, TNS, dynamic, leakage, displacement, and placement legality.
3. Escalate only if ERC improves and timing/power do not regress materially.
4. Roll back on new source-boundary risk, check_placement failure, NonSfinal regression, or added overflow.

Quality Bar

• Never insert boundary-output buffers after sequential, ICG, latch, SRAM, macro, or top-port outputs.
• Do not use broad buffering on sparse giant fanout nets unless boundary filtering proves they are safe internal nets.
• Record every inserted/removed buffer and touched net for audit.