Normalize relocations, addresses, and operands before semantic scoring

[haasonsaas]

Context

The semantic feature model is intended to survive recompilation, but address and operand normalization is implicit in mnemonic-family tokenization. Relocations, literal pools, compiler-generated constants, and register-allocation changes need an explicit, testable normalization contract.

Proposal

Add a canonical operand representation that separates opcode family, operand role, relocation/reference kind, and literal bucket. Exclude raw file/virtual addresses and make the normalization version part of the semantic extractor identity.

Acceptance criteria

• Equivalent functions loaded at different image bases produce identical normalized features.
• Relocations and literal addresses cannot dominate similarity scores.
• Large constant tables and string references are represented by bounded categories.
• Fixtures cover GCC, Clang, MSVC, stripped binaries, and different link layouts.

References

• function feature extraction
• semantic feature design