mr-fang-vlsi/backside-routing-realization-specialist

Backside Routing Realization Specialist

Use this skill when the key question is not "is backside theoretically useful?" but "why is backside routing not being physically/electrically realized, and how do we build, fix, and evolve our own controlled rerouter to test it?"

When to use

Use this skill when:

1. targeted backside reroute reports nonzero selected nets but final DEF shows no BM2/BM1 occupancy,
2. manual DEF patching is being used to test physical realizability,
3. OpenROAD/OpenDB must be used to build a local rerouter or wire rewriter,
4. the project needs a diagnostic bridge between theory-ranked nets and physically realized backside routing,
5. selection bookkeeping must be separated from real route realization.

Scope boundary

This skill owns:

• routing-realization diagnosis for backside signal paths,
• local reroute/patch implementation strategy,
• OpenROAD/OpenDB-backed net wire rewriting and rerouter bring-up,
• understanding the structure of our own routing-related code and patch/replay stack,
• directly modifying and repairing our own rerouter, patch generator, DEF bridge, and related debug tools,
• deciding whether current blockers are in route realization, wire encoding, extraction friendliness, or access path construction.

It does not own:

• workflow ownership or batch orchestration; use it as an expert skill under the active workflow owner,
• the BSPDN physical contract itself without bspdn-physical-contract-auditor,
• PDN sufficiency judgments without bspdn-pdn-sufficiency-evaluator,
• final backside benefit attribution without backside-benefit-attribution-evaluator,
• full batch orchestration.

Core questions

1. Are target nets actually leaving the frontside and occupying BM2/BM1, or is the flow only reporting bookkeeping success?
2. Is the current blocker in route realization, access construction, DEF/OpenDB encoding, or extraction compatibility?
3. Can a local rerouter or wire-rewrite engine physically realize the intended BM2/BM1 -> M0 -> M1 path?
4. Does an OpenROAD/OpenDB-backed patch path provide a more stable realization vehicle than raw DEF text patching?
5. Which part of our own code stack should change: selector, route-spec generation, wire encoder, DEF sanitization, replay contract, or extraction path?

Code ownership expectation

This is not a read-only analysis skill.

When the request requires it, this skill should:

1. inspect our local routing-related code structure,
2. identify the minimum implementation layer that should change,
3. patch the code directly,
4. leave behind a reproducible route spec / patch manifest / script artifact,
5. explain whether the fix belongs in:
   • openroad_backside_rerouter.py
   • sanitize_def_for_openroad.py
   • manual_backside_patch_oracle.py
   • related replay / extraction helpers

The default expectation is "can read, can repair, can extend", not just "can diagnose."

Expected outputs

Emit the smallest useful routing-realization package:

1. *.results.tsv
2. *.conclusion.md
3. optional *.experience_delta.md
4. when implementation is involved, a concrete tool/script artifact and a reproducible route spec or patch manifest

Hard rules

1. Do not treat routed_signal_net_count > 0 as proof of backside realization unless final geometry or extracted connectivity shows real BM2/BM1 use.
2. Keep M0-only separate from confirmed backside success.
3. When manual patching is used, distinguish:
   • geometry preserved in DEF,
   • connectivity valid,
   • extraction valid,
   • electrical benefit measured.
4. Prefer tool-backed route encoding (OpenDB/dbWireEncoder) over repeated raw DEF text surgery once DEF syntax itself becomes a blocker.
5. When modifying our own rerouter stack, prefer the smallest stable implementation layer and avoid copying full industrial router logic unless the lighter wire-rewrite / local-spec / incremental realization path has already been falsified.

Operational references

1. Load references/background-knowledge-links.md for the current routing-realization knowledge contract and source anchors.
2. Load references/update-mechanism.md when deciding whether routing-realization lessons should refresh KB gates, local rerouter assumptions, or the skill's references.