Context
Function features currently include mnemonic-family tokens, instruction mix, an estimated block count, and call/callee degree. The signature does not encode a canonical representation of branch topology.
Proposal
Add bounded CFG-shape features such as normalized block degree histograms, branch-depth buckets, loop indicators, and a canonical graph sketch. Keep them versioned and separately inspectable so the model can be evaluated without obscuring the existing token score.
Acceptance criteria
- CFG features are invariant to file addresses and relocation layout.
- Straight-line, branching, looping, and compiler-reordered fixtures are covered.
- The model records feature-version and component scores in evidence.
- CFG extraction has explicit limits for malformed control flow.
References
Context
Function features currently include mnemonic-family tokens, instruction mix, an estimated block count, and call/callee degree. The signature does not encode a canonical representation of branch topology.
Proposal
Add bounded CFG-shape features such as normalized block degree histograms, branch-depth buckets, loop indicators, and a canonical graph sketch. Keep them versioned and separately inspectable so the model can be evaluated without obscuring the existing token score.
Acceptance criteria
References