[agorevski]

Summary

• Align training export with inference per-function TAC; harden selector provenance, split/gate integrity and metric labels.
• Restore compiler source-mapped opcode reachability in parameterized functions using bounded stack-context CFG traversal.
• Add compiler-based TAC and full-contract runtime identity control benchmarks.

Measured evidence

• Source-mapped TAC witness coverage: 18/36 to 36/36 across six solc 0.8.20 behavior families (structural coverage only).
• In a bounded 2,000-source exact-token audit, usable unique-body rows: 409 to 758; 8k overlength rejects: 618 to 276. Generated datasets/tokenizer are local and not committed.
• Runtime controls: 61/121 checked, with 30 exact equals and 31 exact unequals; 60 context-limited cases correctly unchecked. Exact byte identity is not behavioral equivalence.

Integration and limits

• No trained-model quality gain is claimed; model weights and clean full-run baselines are absent here. The remaining 26.7% long-input tail needs work.
• Main advanced concurrently and git merge-tree reports conflicts in core files. This PR must be reviewed and reconciled before merging; do not force-push main.